hi friends,

this one is for founders and exec teams. if you’ve signed any AI contract, you know the pattern: the tool comes with a dashboard, the dashboard has one big flattering number, and that number slides straight into planning meetings it was never built for.

i want to show you what to do about that, using the version i know best: AI in customer experience.

your team reports “60% automation.” someone does the obvious math: the AI now carries 60% of the CX workload, and the headcount plan, the BPO renewal, and the Q4 temp ramp all get built on that. that math is off by about 2x, and i’ve watched it produce staffing plans that fell apart in the first busy week.

Sheena, our head of deployment (shoutout🥳!), has been building the honest version of this forecast from scratch for individual brands. she figured out everything the dashboards skip by staffing real brands through it. so we turned her work into something you can use, because we want to share how we actually advise brands on this. we built it in Claude Code.

why the dashboard number can’t answer this

automation rate counts messages. it goes up every time the AI writes anything, including an internal note on a conversation your agent fully handled. staffing runs on containment: the share of conversations where no human ever touched the ticket. those are the only ones that take work off your team.

there’s also a middle: conversations the AI starts and a human finishes, with the AI gathering the context that makes the human faster. and a quiet shift underneath all of it: the AI absorbs the easy work first, so what’s left for your team is the complex work, and the average conversation your people handle gets longer and harder, by design. a real staffing plan knows which work is 100% AI, which is 100% human, and how that middle behaves, because it moves handle times on both sides.

a brand can run 60% automation and 25% containment at the same time. both true. with the example brand in the app, “60% automation” sounds like the AI covering 11 agents’ worth of work. the containment math says 6. and the number to staff on is 4. why two numbers? hold that thought.

question 1: how many conversations do we actually handle?

total monthly conversations, all channels, on a trailing 12-month average (a shorter window bakes your peak season into the baseline and double-counts it). plus a holiday-peak setting, because Q4 exists.

question 2: what share can the AI even see?

if the AI covers email and chat but a quarter of your volume is phone and social, that quarter stays 100% human no matter how good the AI is. a 70% automatable share on 75% visibility is really 70% × 75%.

to be fair, most forecasts get the rollout part right. what they miss is that topic mix looks different on every channel, and that live channels run on their own math entirely: arrival patterns, occupancy, shrinkage. in the model, rollouts happen channel by channel like they do in real life: SMS goes live in month 4 and you can watch coverage step from 76% to 83%.

question 3: what are we allowed to automate?

my favorite input, because it puts the responsibility on your policies, not on the AI. for each conversation type: what share is the AI allowed to fully resolve? your refund thresholds, escalation rules, and external claims systems cap it, and no model upgrade changes that. ask whoever owns CX; their answer is the ceiling, and it’s a business decision you already made.

some ceilings sit near zero on purpose. the example includes damage claims that have to be logged in an external quality system: the AI responds, gathers photos, preps the case, and a human always closes it. high automation, low containment, working exactly as designed.

two more things the model handles for you. handle time is per conversation type, because the AI resolves the short conversations first (run this on a blended average and you overstate the answer by close to 20%). and journeys ramp: about 30% of ceiling in month one, about 90% by month six. a forecast at full value from day one means someone’s selling.

and the ceiling is not set once. building an integration in february that gives the AI access to more data? changing your tech stack in june? letting the AI refund higher-value orders in september? every planned change moves the ceiling. the best plans map that roadmap into the forecast, not just the volume.

two numbers, never one

containment gets buffered down for safety. staffing gets buffered up. one number cannot do both jobs.

so the model hands you two. the staffing number runs on a cautious case and goes in the schedule and the BPO commitment, because if the forecast is wrong, your customers feel it first. the value number runs on the expected case, and that one goes in the board deck. the gap between them is printed right on the output, labeled the planning risk window. pretending it’s zero is how staffing plans blow up.

the app also tells you how confident it is: about 60% for a young deployment, rising with history. i’d rather show you 60% confidence with the math visible than a fake-precise number.

what breaks this forecast

two things, and the second is the one almost every forecast gets wrong.

the first is volume shocks. one deployment we worked on saw a conversation type go from 4 to 285 conversations in one week, with zero AI coverage on it. nobody did anything wrong; the volume showed up where the AI wasn’t. the app ranks your conversation types by exactly that exposure, so you know which ones to watch weekly.

the second is mix. topic mix shifts month over month, and once AI is absorbing work, that shift hits staffing much harder than it did when everything was human. your 40% containment in the slow season quietly depends on which topics are doing the volume. if a use case the AI can’t touch spikes alongside your peak rush, containment drops at exactly the moment you needed it to hold. sophisticated forecasts adjust handle time per topic; almost none map each topic’s ceiling against how its volume rises and falls across the year. that mapping is most of what this model is.

one thing only you can decide

this measures the workload the AI absorbs, and it sets no targets for anyone’s team. the brands that get the most out of it decide on purpose where the freed time goes: retention and VIP work, growth covered without new hires, less overtime, a smaller holiday ramp. it’s a much better conversation to have with a real number in hand.

the buildouts

comment below if you want the buildouts or if you’ve built something similar:

• the interactive app: runs in your browser, takes two minutes

• the methodology: every step written out, ready to hand to whoever owns your CX plan

staff for the cautious number. pitch the expected one. keep the gap where everyone can see it. and when any AI dashboard hands you one confident number, ask for the second one.

see you next time.

xx Lisa

hi friends,

i invest in the stock market and i plan to do it even more seriously. so i keep a watchlist of ~20 tech stocks i actually care about, and keeping up with them everyday used to eat real time.

so i built a routine in Claude Code that does it for me.

now every weekday at 12pm ET, i get one slack message. it’s my whole watchlist, biggest movers first, with one plain-english line on why each one is moving. an analyst upgrade, earnings, guidance, a product launch, whatever it is. i read it in about 3 minutes, flag the two or three things worth a closer look, and get on with my day.

here’s exactly what lands in my slack:

the part that actually matters

i prompted it to give me the why, in an easy to understand way.

a list of tickers and percentages is noise. a list of tickers, percentages, and one line explaining each move is a briefing. that one line is the whole difference between “huh” and “oh, i should look at that.”

so i made Claude do two things most stock tickers don’t:

sort by what moved, not alphabetically. biggest movers up top, so my attention goes straight to the action.

explain, don’t just headline. my first version pasted the latest news headline, which was hilariously often about a completely different company. so now it actually reads the news and writes a human sentence about why that stock is moving.

how to build your own (genuinely a weekend-coffee project)

the data: create an account and grab a free api key from finnhub. the free tier is more than enough for a 20-ish stock watchlist.

the brain: a little script pulls each stock’s move plus its recent news. i didn’t write a line of it. i told Claude Code what i wanted and it built the whole thing.

the delivery: a Claude Code routine runs it every weekday at 12pm, skips weekends and market holidays, and DMs me the result on slack. it runs in the cloud, so my laptop can be closed. email works too if that’s more your thing. you just need to give it access to your gmail and password which i’m not sure about.

if “routine” doesn’t ring a bell, i broke down exactly what these are (and how they’re different from scheduled tasks) below:

short version: you brief it once in plain english, connect your tools, pick a schedule, and it runs on its own.

make it yours: swap in your own watchlist at the top. mine’s tech-heavy because that’s what i follow.

total setup is one conversation. it runs in the cloud on a schedule, so there’s no “wait, did i remember to run it.” it just shows up before i do.

steal it this week

pick the one thing you manually check every single day. your numbers, your calendar chaos, your inbox, your watchlist. build a routine that hands it to you before you even ask.

reply and tell me what you built.

xx Lisa

bonjour mes amis,

quick question. how many places did you write down a to-do this week?

a sticky note. the notes app. a notebook you swore you’d actually keep this time. a slack message to yourself. the back of a receipt at dinner.

i did this for years. and here’s the thing nobody tells you. a scattered system isn’t a system. every place you capture something that isn’t connected to anything is just a place for it to die. and the stuff that dies is never the small stuff. it’s the thing you promised someone in a meeting. the intro you said you’d make. the follow-up that would’ve actually moved the deal.

so i stopped trying to be more disciplined about where i write things down. (spoiler: discipline was never the problem can u tell) i changed two things instead.

one way in.

i capture everything by talking now. i use wispr flow and dictate straight into claude code. no separate app, no typing, no “i’ll log it later” (we both know later never comes). if i can say it, it’s captured. two seconds. usually from my phone.

but capturing isn’t the win. you can have the world’s neatest pile of notes and still have no clue what’s actually on your plate. you need something that turns raw capture into one clear view, and keeps it there without you. that’s where this next part comes in.

i’m talking about claude code routines.

okay, what are they actually?

let me back up, because if you’re not living in AI tools all day, things get confusing fast. there are two cousins here - scheduled tasks (claude cowork) and routines (claude code) - and they come from two different claude products. same core idea: you write instructions in plain english, you connect the apps you already use (gmail, calendar, slack, your CRM, your notes), you pick how often it runs, and then it runs on its own, in the cloud (your laptop can be closed), and it messages you when it’s done. no code. you’re basically briefing a new hire and handing them a calendar.

here’s the difference between the two.

scheduled tasks, in claude cowork.

cowork is the chat-style app. a scheduled task is recurring work that comes to you finished. it does the thinking and hands you the output: a recap, a brief, a draft sitting in slack that you just review and send. (i wrote a whole issue on these here →

if you’ve never set one up, genuinely start there. it’s the easiest win in AI right now.)

routines, in claude code.

claude code is a workspace where claude can actually build and run things, not just chat. it can touch files, update systems, make real software. a routine is the same setup (instructions + schedule + your tools), but instead of only handing you a doc, it can do the thing and keep your systems updated. file the action item. update the page. keep your personal OS current.

the one-line version:

- scheduled task (cowork) → recurring work, delivered to you. start here.

- routine (claude code) → recurring work that runs on your systems and keeps them alive.

one hands you the report. the other runs the office while you’re out of the room.

the three i run every week.

1. the morning one.

every day at 7am, before i open a single tab, a routine reads my calendar and sends me one message: a line to get my head right, my top 2-3 priorities, and a clean read of my day. that’s it. i start oriented instead of reactive, which if you’re a founder, you know is the whole game.

how it works: connected to my google calendar, runs at 7am, posts to slack.

2. the meeting one. (this is so good)

this is the one that changed everything for me. after my meetings, a routine reads the notes, pulls out only the things i committed to (the follow-ups, the intros, the decisions i own) and drops them straight onto my calendar, scheduled by how urgent they are. i don’t transcribe action items anymore. i don’t try to remember them. nothing i promised in a room falls through, because it’s not living in my head or on a sticky note. it’s just there.

how it works: connected to my meeting notes (granola + attention) and my calendar, runs every evening, books each commitment as a private block.

3. the writing one.

every sunday a routine scans the week’s AI news and sends me a short brief, written as angles i could actually write about. “what do i write this week” turned into “which of these.”

how it works: searches the web + x + youtube, posts the brief to slack on sundays.

the actual insight.

the goal was never a tidier notebook. it was full visibility with zero upkeep. capture by voice so nothing’s lost on the way in. let the routines turn it into one view so nothing’s lost once it’s there. that’s the whole difference between having notes and actually trusting your system. one of those lets you sleep.

steal it this week.

pick your messiest capture habit: mine was meeting follow-ups, by a mile. for one week, capture it one way only, by voice. then set up one scheduled task in cowork to turn those captures into something you can actually see. just one. that’s the whole assignment.

and on friday:

none of these live on their own. they all feed one place. friday, at the claude code live workshop, i’m going to show you exactly how it connects. we’ll build the personal OS that ties your priorities, your action items, and your plans into one thing. live. in claude code. RSVP here.

come build it with me.

xx Lisa

p.s. want the exact setup for the meeting routine? comment below and i’ll send it to you.

if you’re a founder, you know this one. you own the product, the vision, the brand, and the picture of what “great” actually looks like. so everything routes back to you for the final call.

at Siena that meant every piece of content ran through andrei and i. so whatever the team wrote, my co-founder or i were the last reviewer, every time. product content - Andrei. marketing, social, customer content - Lisa. all of it ended up on our desk for the final pass.

that’s a bottleneck. the fix isn’t to review faster. it’s to put what’s in your head into something that does the first 90% for you.

that’s what skills do. i built twelve of them last month. now a launch goes from “this is shipping” to a full content package across every channel, all sounding like us, without me writing each piece from scratch. here’s how to build your own.

what a skill is

a skill is a set of instructions you write once that teaches AI to do a specific task the way you’d do it. the structure, the voice, what to include, what to skip, how to format the output.

it’s onboarding a new hire. you explain how to write a blog post once, hand it over, and from then on they just know. except this hire never forgets and works in seconds.

skills aren’t new. they’re just underutilized. almost nobody i talk to uses them properly. consider this your reminder.

start from the gap, not the tool

most people ask “what skills should i build?” wrong question.

ask: what do i do over and over that i’m sick of doing from scratch?

for me it was every repetitive content task. so each one became a skill. for you it might be weekly reports, a certain customer email, turning call notes into a follow-up. start with the task you’re already tired of.

how to build one

1. pick one repetitive task.

2. pull real examples of you doing it well. past posts, past emails, whatever you have.

3. tell Claude what you want, give it the examples, ask it to write the skill.

4. test it on a real task.

5. iterate.

step five is the whole game.

the iteration is the point

your first skill will be mediocre. that’s normal. a first draft is a junior hire’s first attempt: 60% there, tried their best.

you have two choices.

quit, fix the output by hand, and decide skills don’t work. or treat it like a real employee: give feedback. “this opening is too corporate. you missed the angle. here’s how i’d phrase it.” run it again. correct it again. 60% becomes 80% becomes 95%.

the people who quit at the first bad output are the ones who say AI isn’t useful. the people who push through get hours back every week.

building a skill is easy. you can write one in minutes. the patience to refine it until you’re 95% happy is the hard part, and the part that matters.

i said 95%, not 100%, on purpose. the last 5 to 10% is your judgment and taste, and that’s still better than the AI. i still read every output before it ships. the skill takes me from blank page to nearly done. i add the cherry on top. that’s the shift: from “create from scratch” to “review and refine.” completely different amounts of work.

where they live

i built and tested everything in Claude. then i moved the skills into Siena so the whole team can use them inside the platform we already work in (we dogfood everything we build). a skill is a shareable file. when our launch motion changes, i update one doc and everyone’s working off the new version.

next i’m putting the repetitive ones on autopilot with scheduled tasks so they run on their own.

what separates a good skill from a useless one (read this twice)

here’s what actually matters. most people write a skill like a vague job description and wonder why the output is generic. the good ones share a few things:

• give it a real point of view, not just a task. a weak skill says “write a blog post.” a strong one says what a good blog post does, what it avoids, and what makes ours different from everyone else’s. you’re encoding taste, not steps. mine literally says things like “verb-led headlines, no ‘meet your X’ framing, short declarative sentences, no unearned adjectives.” that’s a point of view, not a task.

• point it at the masters and the brands you admire. tell the skill whose work to learn from. naming real authors, experts, or companies gives the AI a quality bar it already understands from training. mine references authors like Hemingway for compression and Ogilvy for headlines, plus the brands whose storytelling i actually want to sound like: Linear, Anthropic, Apple. "write like Ogilvy" or "match Linear's product voice" carries more instruction in a few words than a paragraph of me describing what good looks like.

• show it examples, don’t just describe. the fastest way to teach voice is to paste in three things you wrote that you’re proud of. the AI learns more from one real example than from ten adjectives about your “tone.” i don’t tell my skill “write in a casual, authentic voice.” i paste in three of my best linkedin posts and say “this is the bar.” it figures out the voice from the work.

• build in a kill list. half of getting good output is telling the AI what never to do: the corporate words you hate, the formats you don’t use, the clichés that make you cringe. an explicit list of banned moves does more for quality than any amount of positive instruction. mine bans em dashes, the “not X, it’s Y” construction, words like “leverage” and “utilize,” and any sentence that opens with “in today’s fast-paced world.” the kill list does more than any style guidance.

  

• make it ask before it writes. my best skills don’t just run. they ask me for the few things they need first: the angle, the audience, the specifics. a skill that guesses produces slop. a skill that asks produces something usable. my blog skill won’t write a word until it knows the customer, the metrics, and which products they used. it interviews me first, then writes.

• tell it to cut. the highest-leverage instruction in any of my skills is some version of “now remove 20% of this.” AI overwrites by default. forcing a cut at the end is what makes the output feel tight instead of bloated. every one of my skills ends with a step that says “delete 20%, if it can be said in fewer words it must be.” that one line is the difference between tight and bloated.

none of this is complicated. but it’s the difference between a skill you use every week and one you build once and abandon.

pro tip

build one skill to run the others. once you have a handful, add an orchestrator: one skill that maps every deliverable to the skill that handles it. i say "tier 1 launch, [product name]" and it tells me what to produce and which skill writes each piece. the key is that every skill reads the editorial one first, so ten channels still sound like one voice.

if you don’t know where to start and need some inspo, check out anthropic’s and its partners’ directory of skills and plugins.

hope this helps! drop a comment if you want actual prompt skills examples.

otherwise, build, test, and don’t stop at the first mediocre output. give it feedback like a new teammate until you’d actually use what it makes.

that’s the secret. not the building. the not giving up.

xx Lisa

hi friends,

email is the channel marketers actually own. it’s also the channel where the design lift is the most underrated bottleneck.

every newsletter you send goes through some version of: write the brief, hand it to a designer, wait, get a draft, give notes, wait again, finally ship to klaviyo, hope it renders on mobile. multiply that by every product drop, every announcement, every customer moment you want to celebrate. it doesn’t scale, and most of the time it doesn’t happen at all.

claude design changes the math.

last week i ran a private workshop for dozens of brand operators on how to build email newsletters in claude design. got so many DMs after asking when’s the next one? can you share the recording? can my team join?

so i decided to document the process here and host another workshop live where anyone can attend. details at the end.

for this newsletter, i picked one of our customers, Create Wellness, as my example. real fan of their creatine + electrolytes drop (lucky enough to try them right after launch, watermelon is the one). strong brand. clear visual identity. exactly the kind of company that ships fast and needs marketing assets to keep up.

if i were running marketing at Create today, here’s how i’d build a launch newsletter.

step 1: build the brand design system

this is the part most marketers skip, and it’s the reason their claude design output looks generic.

without a design system, you’re prompting a stranger. claude design has no idea what your brand looks like, sounds like, or stands for. so it gives you something that looks like every other AI-generated marketing asset on the internet.

with a design system, you’re prompting someone who already works on your team. they know the fonts. they know the colors. they know the voice. they know what a button is supposed to look like.

the right way to do this: if you’re a marketer at Create (or any brand), you have a real brand guide. a figma file with components, type scales, color tokens, official logo files, the works. you upload the .fig directly into claude design, or you connect the github repo where your design tokens live. claude design ingests the real source of truth, and what you build inherits the actual brand.

what i did for this newsletter: i don’t work at Create. i don’t have access to their figma. so i used a workaround for the sake of teaching. i went to their website, hit command-s in chrome, and saved the page as “webpage, HTML only.” that downloads the page with its fonts, images, and css. then i opened claude design, clicked “set up your design system,” uploaded the HTML file into the “fonts, logos, and assets” field, and added a few notes about brand voice (playful, confident, headlines short and punchy, orange is primary).

ten minutes later, claude design had pulled the typography, the color palette, the buttons, the product photography, and a logo. the brand existed as a system inside the tool.

the catch: this is a hack, not the real thing. an HTML scrape gives you a partial picture. the logo claude generated isn’t the actual Create logo. the type might be off by a weight. some assets are approximations rather than source files. good enough for a teaching example. not good enough to actually ship at Create.

if i were running marketing at Create, i’d skip the HTML hack entirely and upload the real brand kit. 30 seconds, 10x better starting point.

step 2: write the prompt in claude chat first

i don’t write prompts directly in claude design. i write them in claude chat.

here’s why. claude chat lets me iterate on the prompt itself. i can upload references (a screenshot of the brand’s homepage, a few linkedin posts from the founder, a photo from the latest activation). i can ask claude to help me structure the prompt, ask me clarifying questions, and tighten the language before i hand it off to design.

claude chat is for thinking. claude design is for making. also you save tokens by doing so.

for the Create newsletter, i opened claude chat and described what i wanted: a launch newsletter for new york city customers, announcing creatine + electrolytes, anchored on an IRL activation with fitness studios across manhattan. i uploaded the homepage screenshot and the linkedin posts from the launch.

i asked claude chat to write me a structured prompt for claude design. four sections: hero with the product, what’s in it, the irl activation (with an interactive map of partner studios), and a CTA. specific instructions on what’s content versus what’s reference. clear voice guidelines.

the prompt is the leverage. a vague prompt gets you AI slop. a specific prompt, with clear references and clear instructions, gets you something close to shippable on the first try.

then i copied the polished prompt and moved it to claude design.

step 3: edit directly on the canvas

most AI design tools force you to live in the prompt window. you see something you don’t like? type a sentence, hit enter, wait, hope.

claude design lets you edit directly on the design.

you can highlight a headline and rewrite it in place. you can swap copy on a section without re-prompting. you can right-click any element and ask for a specific change without writing a five-sentence instruction.

it also has a comment system. you can drop comments anywhere on the canvas, things like “make this section taller,” “swap this image for the activation photo,” “change this button color.” you can queue up ten comments at once and send them all to claude in one batch instead of prompting one at a time. saves tokens. saves time. closer to how you’d actually work with a designer.

the small thing that surprised me most: claude design adds little touches you didn’t ask for. for the Create newsletter, i asked it to replace text labels in the footer with actual social media icons. it did, and threw in a subtle hover animation on each icon. tiny detail. felt premium.

step 4: export to wherever you ship

claude design has six export options: zip, PDF, powerpoint, canva, HTML, or handoff to claude code.

for newsletters, you have three real paths.

path 1: export HTML, paste into klaviyo or salesforce

this is the fastest. you export as HTML, open your email platform, choose the HTML editor, and paste in the code. it renders.

the caveat: email clients strip a lot of HTML. interactive elements, complex animations, and some css won’t survive. test before you send. but for static newsletters with good layout and typography, this works.

path 2: export HTML, drag into figma, polish, then ship

this is what i recommend for any high-stakes send.

there’s a free figma plugin called “HTML to design.” you install it, drag your claude design HTML export into figma, and it converts the entire newsletter into editable figma layers.

once it’s in figma, you can polish anything claude design didn’t quite nail. final typography tweaks. spacing adjustments. opacity. swapping in higher-res imagery. this is the step where your taste makes the newsletter actually shippable.

from figma, you export the final email assets and load them into klaviyo, salesforce or any other email tool.

path 3: handoff to claude code, ship as a hosted page

this is the path if you want the newsletter to also live as a microsite, not just an inbox send. claude code takes the export and turns it into a live page you can host on vercel or netlify. great for launch announcements where you want a permanent URL to link to from social, your inbox, your sales team. i talked about this more in depth in my previous newsletter.

the part most people get wrong

claude design will get you to 95% on autopilot.

the last 5% is everything. it’s the typography choice that makes the whole layout feel intentional. it’s the line of copy that goes from “fine” to “exactly right.” it’s the spacing tweak that makes the section breathe. it’s catching that one image that looks generic and swapping it for one with real personality.

that 5% is the difference between a newsletter that looks like AI made it and one that looks like you made it.

AI doesn’t have the taste you do. the workflow above gets you to the version where your taste matters. that’s the whole point.

see it live

due to high demand, i’m running an open workshop on claude design. we’ll build a newsletter together from scratch, using a real brand. i’ll share the hacks i didn’t cover here, the small prompting moves, the editing shortcuts, the export tweaks, the things that get your output from “decent” to “i can’t believe i made this in minutes”.

it’s free and open to everyone.

sign up here.

what brand do you want me to build a newsletter for live? hit reply and tell me.

xx Lisa

hello beautiful people,

so we’re throwing a launch event on may 13.

it’s called Siena Renaissance. and it’s the first time we’re bringing the whole community together. it’s going to be great!

three weeks ago we wanted a real landing page for it. not just a luma link. something that actually said the word “renaissance” the way i meant it. type that belonged on a gallery wall. a pixelated bust that animated as the page loaded. a team carousel of every Sienaer helping pull this off. a real RSVP page.

and guess what! Claude Design had just launched.

what came out of that weekend was the landing page, the music, the video, and a workflow that’s now how i build everything. here’s the play-by-play.

PS: when i tell you i am glued to my deck because of Claude, i am not kidding.

step 1: i built our brand design system in Claude Design

before i could build the landing page, i needed our brand to actually exist as a system. fonts, colors, typography rules, component styles. the kind of thing a real design team builds in figma over weeks (otherwise u just end up with AI slop).

Claude Design lets you define a full brand system inside a single project. i described what Siena should feel like: classical but digital, cream-on-ink palette, feature deck for headings, sans for body, lots of breathing room, no clutter.

30mins later, the brand system existed. typography defined. colors locked. component patterns saved. every page i built after this inherited the system automatically.

this was the unlock. once the brand existed as a system, every piece of collateral i made next was consistent without me thinking about it.

step 2: i mocked the whole landing page

with the brand system in place, i opened a new project for the event landing page.

i described what i wanted: our designer worked on a beautiful pixelated Renaissance bust that i uploaded, then defined my vision like eight sections from intro to RSVP. then i iterated. change the hero. add a customer logos row. pixel-animate the statue. now make the statue more refined.

about 20 iterations in, the page looked the way i’d imagined it.

Claude Design has a feature called “export handoff.” it bundles the whole project (HTML, CSS, fonts, assets, a README that literally says “CODING AGENTS: READ THIS FIRST”) and you can directly hand if off to Claude Code.

step 3: Claude Code turned it into a live site

Claude code is Claude in a terminal. it reads files, writes files, runs commands. i asked it to implement the page for real. my goal was to be able to push it live.

before going live, i iterated with it the way i iterate with my team:

• “make sure the entire Renaissance word is on one line”

• “add a carousel after ‘who’s in the room’ with five people from the Siena team”

• “make it scroll automatically, slowly”

• “add some easter eggs for the RSVP button’”

• “make the copy more renaissance!”

when i was happy, i asked it to deploy. it told me one of the easiest ways is to push it via Vercel. i pasted the command. Vercel asked six yes/no questions. 8 seconds later the build was done. URL in hand. wooo!

step 4: pushing to Github (the part most people skip, including me at first)

if you’re working solo on a one-off, Vercel is enough. the URL works, the page is live, you’re done.

but the second i wanted my team to access the repo, push their own edits, or set up a custom domain, Vercel alone wasn’t enough. i needed it on Github.

i asked Claude Code to push the project to our company github repo. it walked me through creating a new repo for these types of projects, set up the remote, committed everything, and pushed. then it linked the Github repo back to vercel so every push to main automatically deploys to production.

now anyone on the team can pull the repo, make edits in Claude Code, push, and the live site updates. no more “send me the file” or “is this the latest version.” Github is the source of truth.

if you’re building something you want to keep evolving (and especially if you want a custom domain), don’t skip this step. one-off deploys are fine until they aren’t. (my rookie mistake)

step 5: the loop is where it got interesting

i thought i was done. i kept editing after i shared it with my co-founder.

over the next few hours, i kept opening Claude code and sending it screenshots of my own page with annotations. “make this say that.” “move this here.” “change the section heading from ‘the customer panel’ to ‘the patrons.’” each time it edited the file, i refreshed, decided if i liked it, and either asked for another tweak or asked it to deploy.

the deploy command became reflexive. npx vercel --prod. 8 seconds. done.

a few things that ended up on the page only because the loop was that fast:

hover effects on the RSVP button. “when i hover the RSVP button on the hero, change the text to ‘AI made the page. you make the room.’ two minutes later it was working in production. who doesn’t love easter eggs?!

a team carousel that auto-scrolls. i sent Claude Code each team member’s photo and renamed renaissance title: the curator, the architect, the emissary, the maker, the steward, the patron’s voice, the master of works. i asked for an automatic slow marquee and voila.

copy edits as conversation. every section heading went through 3-4 versions. “the customer panel” became “the panel” became “the patrons.” “the agenda” became “the program.” “the team” became “your hosts” became “the Sienaers.” each one was a sentence to Claude Code, fifteen seconds to deploy, instant feedback on the live URL.

if i’d been working with a designer plus a developer, every one of these would have been an email, a figma comment, a sprint, a back-and-forth. instead they were typed sentences. shipped in seconds. DOPAMINE WAS HIGH.

step 6: the music

i wanted a song for the event video. not stock music. something that felt like a Manhattan rooftop in May at 6pm. golden hour. drinks in hand. people arriving.

i used Suno, described the vibe in detail, and named reference artists: Folamour, Daft Punk, Roy Davis Jr., Crackazat, Kaytranada. it generated 6 versions in a few minutes. i picked the one that hit.

then i needed to edit it down. the original was 3 minutes, i needed 30 seconds. i opened Claude Code again. asked it to write me a quick command using ffmpeg to trim the track. it gave me the command, i pasted it into terminal, the file was trimmed.

step 7: the video

ok so last night at 12am it hit me. the event needed a teaser video. the pixelated bust animating. the wordmark appearing. the music behind it. classical but digital. the whole renaissance feeling, in 30 seconds.

i built the visual in Claude Cesign as a separate page with frame-by-frame animation. then i used QuickTime to screen record it. then i went back to Claude Code and asked it to combine the screen recording with the trimmed Suno song into a single MP4. one command later, the file existed. i dragged it into LinkedIn. it played.

what changed isn’t “AI got smarter.” what changed is the loop.

the pieces of the workflow are now actually compatible with each other. Claude Design holds the brand system. Claude Design exports the bundle. Claude Code reads the bundle. GitHub stores the source of truth. Suno generates audio. Claude Code edits the audio. QuickTime records the screen. Claude Code merges everything. Vercel deploys from GitHub to production. it's one continuous motion.

the unlock isn’t any single tool. it’s that they form a loop you can run inside a single conversation, in real time, with your own taste as the only constraint.

the part that’s underestimated: brand-level marketing collateral is now a solo sport.

i keep thinking about who this is for.

it’s for every founder who wanted a better marketing site but didn’t have a designer on payroll. or maybe they have a designer but want to execute across multiple projects at the same time. it’s for every ops lead who wanted a clean internal portal and got told “that’s a Q3 project.” it’s for every CX lead who wants their help center to look like their brand, not like a zendesk default. it’s for every founder who couldn’t afford a videographer for their launch.

this isn’t AI replacing designers. great designers are now becoming the orchestrators of these AI supertools and they should still own your brand system, your product UI, your marketing sites.

but for the long tail of collateral no one ever had budget for properly, the math has flipped. founders can ship their own marketing the way they already ship their own investor updates and their own pitch decks.

anything you’ve been putting off because “we don’t have a designer for that right now” probably doesn’t need one anymore.

wbu? have you tried Claude Design yet? tell me. i want to see it.

xx Lisa

PS: if you work in ecommerce, come to our event!

let me be honest about something. this is probably the most basic AI use case i could write about.

it’s meeting recaps. not a fancy agent. not a complex workflow. meeting recaps.

but here’s why i’m writing about it: most people i know are still copy-pasting transcripts into Claude and prompting from scratch every single time. or worse, relying on whatever generic summary their AI notetaker spits out. (most AI notetakers are bad)

i use Granola for transcripts, and it’s the best one i’ve found. the summaries are genuinely good. but even with Granola, i was still taking the transcript, pasting it into Claude, and writing a better prompt because i wanted specific things surfaced for specific stakeholders. Granola’s default summary was never quite what my team needed to hear.

so i was doing this manually, every week, for three recurring meetings. sometimes i did it. most weeks i didn’t. and the context just died.

i finally spent two hours building a system that does all of it automatically. and it will probably save me two to three hours per week for as long as these meetings exist.

how it works

i went into claude and pulled the last several transcripts from Granola for each of my recurring meetings: leadership weekly, sales team weekly, product marketing weekly. i studied the patterns across weeks. what topics come up, who attends, what’s useful vs. noise.

then i built a “skill” for each meeting. a skill is a set of instructions that tells claude exactly how to write the update. the structure, the tone, what sections to include, what to skip, how to format for slack.

each skill is different because each meeting is different. my sales update leads with attainment numbers and goes deal by deal. my leadership update covers product, customer health, commercial strategy, people, and ops. my product marketing update is tighter because the meeting is more focused.

i tested each skill on a real transcript, iterated until the output was something i’d actually post, then scheduled it.

the scheduled task

by now we all know cowork has a feature called scheduled tasks. you give it a prompt, a frequency, and it runs automatically with access to all your connected tools.

for each meeting i created a task that fires ~30 minutes after the call ends. the prompt tells claude to pull today’s transcript from Granola (fallback Attention or Gemini), write the update using my skill, and draft it to the right slack channel.

it drafts, not sends. i review it, hit send. 30 seconds vs. 10+ minutes.

once you’re comfortable the output is good, you can remove the review step and let it post directly. the transcript source doesn’t matter. Granola, Attention, Gemini, whatever records your calls. the logic is the same.

this is the gateway drug. here’s what else you can schedule.

meeting recaps are just the most obvious starting point. once you realize that scheduled tasks can run anything claude can do, on a recurring basis, with access to all your connected tools, the list gets long fast.

weekly competitive briefing. every monday morning, pull the latest from competitor websites, job postings, review sites, and social accounts. summarize what changed. post to a slack channel. your team starts the week knowing what’s happening in the market without anyone manually checking.

pipeline snapshot. pull data from HubSpot, cross-reference with meeting notes from the week’s sales calls, and draft a pipeline summary for leadership. no more manually building the update in a spreadsheet.

customer health digest. pull the last week of customer calls from Attention, surface accounts showing risk signals (complaints, competitor mentions, contract pushback), and post a summary to your CS channel. early warning system that runs itself.

hiring updates. pull candidate pipeline from your ATS, summarize where each role stands, flag anything stalled, draft a slack update for the hiring team. nobody asks “where are we on hiring?” in the weekly meeting anymore.

board prep inputs. every month, pull ARR, pipeline, customer metrics, and product milestones from your tools and compile a first draft of your board update. you still write the narrative. the data assembly is done.

the one that’s actually underestimated: your social presence

i know a lot of people have been struggling to be consistent with content. they have interesting things to say but never find the time to write them down.

here’s what i’m building for myself: a scheduled task that pulls the most interesting insights from my conversations each week. client calls, partner meetings, events, internal discussions. it surfaces patterns and themes i might not have noticed, then drafts a linkedin post using a skill i built that knows my writing voice, my editorial preferences, and the topics i write about.

the result is a draft every week that starts from my actual conversations and my actual point of view. not generic AI slop. content that comes from things i said, were said to me, things i learned, things i believe. i just edit and publish.

if you build a strong skill (think of it as your editorial guide: your voice, your style, the type of expertise you write about), and combine it with a scheduled task that feeds it real context from your week, you have a system that helps you stay consistent without it feeling like a chore.

your conversations are already full of content worth sharing. you just need a system that surfaces it.

scheduled tasks are underestimated

most people think of AI as something you chat with when you need help. scheduled tasks flip that. AI comes to you with the work already done.

anything you’ve been putting off because “i don’t have time for that,” anything in your backlog that keeps getting pushed, there’s a good chance it can run on autopilot.

start with one meeting. see how it feels. then look at your week and ask: what else should be running without me?

what have you put on autopilot? reply and tell me. i want to hear what others are building!

see you next time.

xx Lisa

hey so listen… i am not a VP of sales. not a traditional one at least.

i’m a co-founder/president running growth, sales, marketing, partnerships, and community. there is no version of this job where i have time to babysit a pipeline spreadsheet or prep every deal review from scratch.

so i built a system. claude, connected to the tools already running our business.

here’s what that actually looks like.

oh but first cause it’s morning…

okay the stack

eight connectors live inside claude:

• attention AI -- call recordings, transcripts, and a superagent in slack for asking questions across all conversations

• granola -- internal meeting notes and context layer across the entire org

• hubspot -- pipeline and CRM (attention auto-populates fields here based on our call workflows)

• gmail + google calendar -- email history and meeting context

• slack -- basically our entire existance (ugh i know)

• Siena MCP -- our accounts, CX data, and benchmarking

• linear -- product and engineering

• gamma -- for decks and presentations

the unlock isn’t any one tool. it’s that claude pulls from all of them at once, in a single conversation. one tab instead of twenty.

getting up to speed on any deal

before a customer call, i ask claude to pull the full picture: hubspot record, recent emails, last call summary from attention, relevant slack threads, upcoming calendar context. two minutes. sometimes from my phone.

renewals and expansions

one prompt pulls usage data from Siena, contract details from hubspot, open issues from slack, and recent transcripts from attention. output: where we’re strong, where we’re at risk, what the ask should be, what objections to expect.

and then we take it one step further -- we push our QBR data directly into a gamma deck from claude. we walk into every customer QBR with a fully prepared deck, built from live data, not assembled the night before from five different sources.

forecasting and pipeline review

every week i pull all open opps, flag anything without recent activity, and surface deals where close dates have slipped. we also do monthly and quarterly retros led by Andrew, our Sr Manager of Sales, where he uses claude to pull highlights and lowlights across all conversations in that period -- patterns, trends, what to focus on -- combined with our numbers from hubspot dashboards. it keeps the whole team aligned on where we actually stand.

deal reviews

i drop a deal name and ask for the full picture: what we pitched, objections from transcripts, what’s been sent over email, who the stakeholders are, where it stalled. then i ask what to do next.

competitive intelligence

attention captures everything said on a call. i run a sweep of recent transcripts and ask claude to pull every competitor mention, what the prospect said, which features came up. that output goes three places: sales slides, marketing campaigns, and directly into website copy. what used to require a dedicated PMM is now a tuesday afternoon.

win/loss analysis

end of month: sweep closed-won and closed-lost in hubspot, cross-reference with transcripts. patterns: what do we win on, what kills us, who do we lose to most and why. 30 minutes with a good prompt.

board meeting prep

this used to involve multiple stakeholders spending hours pulling data from different places. now i run one prompt that surfaces highlights, lowlights, trends, goals, and a forecast for the next quarter -- numbers included. the narrative still needs a human. the brief doesn’t.

performance reviews

pull all deals an AE touched in a quarter, cross-reference with attention call scorecards, flag patterns in how they handle objections. conversations become grounded in actual data.

meeting prep

every customer call, every prospect meeting. claude reads the calendar invite, pulls context from every connected system, surfaces what i should focus on and what to watch for. granola is especially useful here for internal meetings -- the context it builds across the org means i’m never walking in cold.

product feedback loop

my favorite one. customers drop product feedback on calls, usually in passing. a weekly prompt sweeps those transcripts, extracts product mentions, and sends them straight into linear. feedback that used to die in a recording now reaches engineering.

Siena MCP as an intelligence layer

we build Siena, so we connected our own platform to claude too. that gives me fast access to account data, usage signals, and benchmarks without pulling a manual report or asking someone to dig it up. CX, product, and leadership are all working from the same source of truth. fewer “can you pull that for me” requests, fewer one-off slack messages asking someone to dig up account data.

thought leadership at 1am

some of my best linkedin posts come from things customers say on calls. a frustration i hadn’t heard before. a way someone described their problem that just clicks. a question that keeps coming up across accounts. moments that stood out. it’s the fastest way i’ve found to write something worth saying.

because i’ve given claude access to all of these tools, it’s not just surfacing information but drafting and taking action. a follow-up email. a new linear ticket from a customer call. a slack message. a gamma deck.

for now, i double check everything before it gets pushed (bit OCD one would say). that’s a personal choice, not a limitation of the tool.

but my next step is to give it more autonomy. claude code or cowork works even faster. it’s really up to you how much you want to micromanage your AI agent.

tl;dr

we made one decision that changed how the whole org operates: treat every conversation, sales, CX, internal, as structured data. the result is no more tabs. no more asking five people for context before a meeting. no more losing product feedback in a call recording from six weeks ago.

as a founder running sales, the biggest shift is that your next step is right there.

how are you running your sales org?

yours truly,

Lisa xx

hi friends,

confession time: i’ve been itching to share more video—customer wins, AI insights, founder mishaps—you name it. i started looking on upwork for a freelance editor… then i stumbled on the silicon valley girl podcast show notes (yes, that rabbit hole again) and discovered OpusClip and Descript. so i got on the Opus free trial.

i uploaded our long-form customer interview with Tanner Nelson from Momentous (best magnesium l-threonate ever btw)—and in under four minutes opus spit out 27 perfectly captioned shorts. mind. blown.

here’s why i’m keeping it, and how you can steal the playbook.

why opus clip matters

• zero-to-clips in minutes paste a youtube / zoom link or drop an mp4 → opus finds the hooks, adds dynamic captions & emojis, and exports for tiktok / reels / shorts. you get 5–10 (or in my case, 27) ready-to-post clips faster than coffee brews. even if my primary channels are linkedin and x, the formats still work.

• built-in virality score opus ranks each cut against a giant library of viral videos, so you post the highest-potential clips first.

• hands-off branding upload your logo + hex codes once. every caption bar, emoji, and thumbnail stays on-brand.

• scheduling you can use the calendar feature to schedule your posts in advance.

how to start

1. sign up at opusclip.com (free tier gives you 60 minutes).

2. upload or link a long video

3. select “talking head + captions, 9:16.” keep it simple.

4. hit generate and watch the magic.

5. download & post.

it’s not always perfect so you can edit the copy, use AI features like adding an AI hook or a B-roll to polish it.

how i plan on using it

• customer stories → Tanner’s Momentous interview now lives as 27 snackable shorts for LinkedIn, and X.

• product-launch recaps → 30-min launch call becomes 8 reels: feature demo, customer quote, behind-the-scenes.

• webinar highlights → slice our webinars into small actionable chunks.

• founder journey learnings → bite-size clips on building siena, news, learnings.

• AI education → short explainer vids breaking down big concepts, benchmarks, prompting tricks, or model comparisons—perfect for keeping our community up-to-speed.

my goal is to have on-brand clips that keep content flowing.

my pov is that AI isn’t replacing creativity, but it obliterates busywork. the hours we used to spend in other tools can now be redirected to more important stuff.

my challenge to you: run your next recording through opus, post one short within 24h, tag me—I want to see!

next week: generative search (geo) and how to make your brand “AI-discoverable.” stay tuned.

xx Lisa

hey you!

this year has been a lot to keep up with, right? AI launches happening every day.

i run the business side of Siena aka sales, partnerships, GTM. part of my job is to make sure we adopt the latest tools, build agentic workflows for our team, and stay ahead. i work in and with AI every day and i still couldn’t tell you which of these four tools to pick.

until i started testing all of them.

friends of mine have been buying mac minis to run openclaw (the viral agent formerly known as clawdbot) on dedicated machines. my own tool stack right now includes the claude desktop app (heavy on projects and connectors), claude code, perplexity comet and i just started testing cowork and perplexity computer.

at some point i stopped and asked myself: what exactly should i be using? and more importantly, what should i be recommending to operators, CX leaders, and founders who read this newsletter?

so i did what i always do. i went deep. i ran research across perplexity, chatgpt, and claude, compared the notes, and tested the tools myself.

here’s what i found and who should use what.

claude code - the one worth the learning curve

claude code lives in your terminal. you type what you want in plain english, and it writes code, edits files, builds tools, runs scripts, and commits to github. it has a million-token context window, which means it can hold your entire project in its head at once.

i know - “terminal” sounds intimidating. i hardly ever opened one before this. our team uses it heavily so i challenged myself to try it, and here’s what i found: the discomfort lasts about 15 minutes. after that, you’re just typing instructions in plain english. the interface is different. the experience isn’t that different from chatting with AI. and i use claude to help me use claude code.

what makes claude code worth it is what you can build with it. and for founders running 7- and 8-figure+ brands, the use cases go way beyond code.

you’re trying to build a 12-month financial forecast that pulls from your shopify revenue data, ad spend from meta and google, and COGS from your 3PL. claude code can build a model that connects to those data sources and updates automatically - no more copy-pasting into spreadsheets every month.

you want to analyze your last 6 months of unit economics by SKU - gross margin, CAC, LTV, return rate - and flag which products are actually making you money after all costs. claude code can process that entire dataset and output a clean dashboard.

your ops team is spending 10 hours a week reconciling inventory across shopify, amazon, and your warehouse. claude code can build a script that runs that reconciliation daily and alerts you to discrepancies before they become oversells.

you need a working prototype of a customer-facing returns portal or a loyalty program calculator before committing engineering resources. claude code can build a functional version in an afternoon.

the limitations are real though. token consumption is significant - a typical multi-file task burns 40,000-80,000 tokens per session. rate limits can pause you mid-task. and as the context window fills up, it can drift from your original instructions. always review the output.

who it’s for: founders who want to build operational tools, automate data workflows, prototype fast and consolidate as much as possible (context window is key here). if you’re a brand founder or head of ecommerce choosing between hiring and AI - this is the tool that tips that equation.

what it costs: $20-$200/month depending on your usage tier.

claude cowork - the fastest way to start if you’re non-technical

cowork is a tab inside the claude desktop app. you point it at a folder, tell it what you want, and it creates spreadsheets, cleans data, organizes files, and builds documents. it has a lot of connectors like google drive, gmail, slack, notion, or salesforce.

this is where CX leaders should start. the workflows it handles well are exactly the ones eating your team’s time:

you need to analyze 6 months of CSAT data broken down by agent, channel, and ticket category to build a case for headcount or restructuring. cowork can process the export and generate the analysis doc.

you’re synthesizing customer feedback from zendesk, social, and reviews into a single prioritized report for your product team. cowork can pull it together across channels.

you need to clean a messy product data export before a migration or catalog update. cowork handles that without you touching a single formula.

it also works well for competitive pricing analyses, organizing shared drives, and building slide decks from raw data.

the limitations: it burns through your daily usage fast - 3-5 serious tasks on the $20 plan and you might hit the ceiling. connectors fail about 25% of the time. it drifts on longer tasks - you’ll get 90% of what you asked for. and there’s a well-documented incident where it permanently deleted 15,000 photos from someone’s computer (they had icloud backups). always back up before pointing it at important folders!

who it’s for: CX leaders, ops managers, anyone who works with files and data and wants to save hours on repetitive analysis and document creation. no terminal, no setup friction - you can be productive in minutes.

what it costs: included with any paid claude plan ($20-$200/month).

perplexity computer - i’m testing it right now

i’m not going to pretend i have a full opinion on this yet. i started testing perplexity computer this week and i want to give it a real evaluation before i tell you what i think.

what i can tell you so far: seems to be the most ambitious of the four. It orchestrates 19+ frontier AI models like Claude Opus 4.6 handles core reasoning, Gemini tackles extensive research, ChatGPT 5.2 manages long-context recall, Grok handles lightweight tasks, and specialized models generate images and video. the system automatically routes each subtask to the best-suited model, creates sub-agents, and runs them in parallel.

it’s $200/month. that’s the only tier available right now.

i’ll go deep on this one in the next edition with real workflows and honest results. so excited about it. stay tuned!

openclaw (formerly clawdbot) - the one everyone’s losing their minds over

this is the viral one. a free, open-source AI assistant you host on your own machine. you message it through whatsapp, slack, or telegram, and it reads your email, manages your calendar, browses the web, and executes real tasks autonomously. one user’s agent negotiated $4,200 off a car purchase over email while he slept.

it got 145,000+ github stars in weeks. friends of mine are buying dedicated mac minis just to run it 24/7. the hype is real.

so is the risk. it gives an AI agent shell access to your computer, your email, your API keys. in one community test, an agent deleted all emails after receiving a message with hidden instructions. security researchers found 512 vulnerabilities in the codebase. the creator just left to join openai.

if you’re a technical founder who wants to understand where personal AI agents are heading, it’s a fascinating sandbox. for everyone else - watch from a distance or test it just don’t give it your work credentials.

what it costs: free (you pay for AI model API costs, or nothing if you run local models). setup takes 1-2 hours of command line work.

what’s still hype

none of these tools reliably handle novel, ambiguous, or high-stakes decisions without human oversight. claude code struggles with unusual architectural patterns. cowork can stall mid-task on complex workflows. perplexity computer’s orchestration sometimes produces mediocre results on poorly-defined prompts. openclaw’s security posture has led to actual account bans and data exposure.

if someone tells you they’re running their business on autopilot with any of these tools, they’re either exaggerating or about to learn an expensive lesson.

the pattern to watch

the real shift here isn’t any single tool. it’s the transition from AI as a chat partner to AI as a background worker.

every tool on this list is trying to move from a conversation interface to a delegation interface - where you describe outcomes and the AI figures out the steps. that model is early but accelerating fast.

the companies that learn to delegate effectively to these tools will have a meaningful operational advantage over those still copy-pasting from chatgpt. that’s the gap that’s opening up right now.

so what should you actually do?

you don’t need all of these. you can pick one and see if it works based on your role and what you’re responsible for.

if you’re a brand founder or head of ecommerce: start with claude code. build one automated workflow - a financial model, an inventory reconciliation, a weekly metrics dashboard. the terminal feels unfamiliar for 15 minutes and then it’s just typing. alternatively, if your biggest need is research and competitive analysis at scale, perplexity computer is worth trying.

if you’re a CX leader or ops manager: start with cowork. it handles the exact workflows eating your time - CSAT analysis, feedback synthesis, data cleanup, reporting. no terminal, no setup friction. you’ll be productive in minutes.

if you’re a CTO or technical founder: you probably already know about claude code. look at perplexity computer for research-heavy projects and keep an eye on openclaw for what it signals about the future of personal AI agents.

if you’re curious and have time to experiment: try more than one. i’m using claude code, cowork, and perplexity computer for different things. the tools can be complementary, not competing. the key is matching the right tool to the right workflow. ideally i want to consolidate as much as possible into a single one.

the pattern across everything i tested is the same: the demos look incredible, the reality requires oversight. nothing here replaces your judgment. but the right tool, matched to your actual workflow, saves real hours every week.

that’s where we are. not magic. not useless. somewhere in between and moving fast!

hope this brought some clarity. let me know what you’re testing and building!

xx

Lisa

bonjour friends,

ever cringed at AI-generated content because it felt... off-brand? same. the truth is, if your AI doesn’t sound like you, it’s probably because you haven’t trained it properly yet.

AI tools like ChatGPT and Claude can be incredible extensions of your team but only if you teach them how your brand actually sounds. the good news? training your AI isn’t rocket science. think of it like onboarding a new team member: show them examples, explain your preferences, give clear feedback, and watch the magic happen.

here’s exactly how we trained our own AI at Siena AI to nail our unique brand voice:

step 1: start with your brand essentials

gather all your core brand materials in one place:

• your writing style (highly encourage you to create one, comment or DM me if you want my template)

• your social media brand voice guide

• examples of your best-performing content, newsletters, and posts—everything that captures your true vibe and personality

think of these as your AI’s onboarding docs.

step 2: build your AI’s “brand bible”

drop these guidelines and your favorite writing examples directly into ChatGPT (i use GPT-4.5) or Claude (Sonnet 4 is fantastic) in a dedicated “project”.

clearly explain why you chose each example:

• “i love the directness and humor in this newsletter intro”

• “the conversational tone here feels exactly right”

• “this social post perfectly captures our data-driven, customer-first approach”

AI learns exactly like humans: from examples and detailed feedback. the more specific you are, the better it gets.

step 3: experiment, iterate, and edit

prompt the AI to rewrite existing content (newsletters, social posts, emails) based on the examples you provided. your prompts might look like:

• “rewrite this newsletter intro with the same conversational and slightly quirky tone as our last edition.”

• “make this product update less formal, inject some subtle humor, and remove jargon.”

always leave time for editing rounds. remember: AI supercharges your ideas, it doesn’t replace your judgment.

step 4: set clear brand boundaries

give your AI specific guardrails so it stays true to your brand:

• do: be conversational, friendly, clear, empathic, and occasionally funny.

• don’t: use jargon, use emojis, sound overly corporate, hypey, or excessively formal.

• always: write naturally, authentically, and in a way that’s easy for everyone to understand.

step 5: involve your team

share your AI brand guidelines with your whole team. everyone should use the same AI training approach to keep your brand voice consistent across every touchpoint.

bonus: create a shared internal document with your team’s favorite AI-generated content pieces and prompts that “just nailed it.” this way, everyone knows exactly what “great” looks like.

some tips

• chatgpt vs claude: try both! claude (sonnet 4) is super conversational and natural, ideal for initial drafts. chatgpt (4.5) is highly reliable and remembers your style over time—perfect for polished final edits.

• feedback loops: AI learns faster with frequent, specific feedback. if something’s off, explain exactly what and why this small habit accelerates AI quality dramatically.

• regular updates: refresh your AI training regularly with new examples as your brand voice evolves.

• use a writer’s checklist: in your prompt, challenge the AI with questions that would make it follow your guidelines strictly.

why this is a big deal

• consistent quality across every team member and communication channel.

• faster, easier content creation without sacrificing authenticity.

• building institutional brand knowledge that outlives team changes, vacation gaps, or busy periods.

bottom line: this upfront training is worth it. invest time now, and watch your AI-generated content become something you genuinely love (instead of something you constantly need to fix).

what’s the hardest part of training AI to sound like your brand? lmk!

bisous,

Lisa

hi friends! i missed you. been heads down into the other 1000 priority items at siena haha!

i wanted to really try this new tool that’s been going viral on X and report back. so last week i set up remotion which is a react framework that lets you create videos programmatically and connected it to claude code.

here’s why i started experimenting.

the old way

you want a 30-second feature demo video. here’s what that looks like: write a brief, send it to a designer or video editor, wait a few days, get a first draft back, give feedback, wait again, get a second version, more feedback, maybe a third round. if you’re lucky you have it in a week. if your video person is busy, two weeks. multiply that by every feature launch, sales asset, and customer story your team needs. it doesn’t scale.

and if you want to change one line of copy or swap a screenshot three months later? you’re starting the cycle over.

where this is heading

with remotion + claude code, the workflow is completely different. i created a claude code project on my macbook and loaded it with everything like product screenshots, pitch decks, notion docs, branding guidelines, our design system from github, font files. basically gave it the full picture of what Siena is, how we talk about it, and how it looks.

then i told claude code to build an animated explainer video for Siena. it pulled from all of that context like the messaging, the visuals, the brand and generated a video composition. text animating in, transitions between sections. all in code, all previewable in the browser in real time.

i connected our design system repo from github so the videos pull our brand fonts, colors, and gradients automatically. no back and forth with design to make it look on brand. it already is.

want to change the copy? update one line. swap a screenshot? drop in a new image and re-render. the whole thing lives in code, so every edit is instant.

here’s one of the videos i built LOL

how to get started

here’s the thing… i don’t code. i don’t know react. i’ve never used remotion before last week.

my entire workflow is sending screenshots to claude and asking “what should i do next.” that’s it. at every single step, claude gave me the exact command to run or the exact prompt to give claude code. here’s how the process went:

1. asked claude if i have node.js installed. it told me to run node -v. done.

2. asked claude how to set up remotion. it gave me one command: npx create-video@latest. when the template options came up, i screenshot them and asked which one to pick. it said Hello World.

   

3. asked where to put product docs so claude code can read them. it told me to create a docs/ folder. asked where to put screenshots and fonts. public/ folder. asked where to put our design system. clone the repo into docs/design-system.

   

4. opened claude code inside the project and gave it my first prompt: read the docs, look at the screenshots, build a Siena explainer video. claude wrote the exact prompt for me.

5. previewed it in the browser. it looked rough. sent a screenshot back to claude, said “this looks messy, help me fix it.” it gave me a polish prompt to pass to claude code. ran it. better.

6. kept going. one prompt at a time. fix the fonts. smooth the transitions. make the cursor feel human. make the typing speed natural. each round, a little closer to something i’d actually ship.

every step of the way, i didn’t need to know what i was doing. i just needed to describe what i wanted and show claude where i was stuck. that’s the workflow.

how to get started in 4 steps

1. install claude code! this is regardless of any tools. it’s a game-changer (will talk about it in my next newsletter, just do it!). docs here.

2. install remotion — watch this video, docs here or just ask claude tbh 😅

3. load your context — create a docs/ folder in your mac and drop in everything: product docs, brand guidelines, decks, screenshots, fonts. the more you give it, the better the output.

4. open claude code and start prompting — one prompt at a time. describe what you want, let it build, preview in the browser, iterate.

the honest limitations

i’m not going to pretend this is plug and play. here’s what i learned:

you need to feed it a lot of context. product docs, brand guidelines, design examples, fonts, decks, notion pages. all of that needs to live in your project so claude code can reference it. the more you give it, the closer the output gets to what you actually want. garbage in, garbage out.

it takes a lot of iterations. the first output looked like a junior designer made it. text spacing was off, transitions were abrupt, screenshots felt like a slideshow. each round of prompting got it closer, but expect to go back and forth.

you’re not getting crazy SaaS motion graphics out of this. our solutions engineer creates polished animated videos that are on another level. remotion + claude code won’t replace that kind of work YET. but for quick product demos, sales enablement, and feature walkthroughs? it’s more than enough.

there’s still a lot of manual editing. prompting gets you 60% there. the remaining 40% is you tweaking timing, adjusting layouts, fixing small details that AI doesn’t nail on the first try.

it works best one prompt at a time. if you throw too much at claude code in a single prompt like redesign everything, change the fonts, add new animations, fix the timing, it gets overwhelmed and the output suffers. break it into focused steps.

and the better your systems are in github, the easier this gets. if your design tokens, brand assets, and product docs are well-organized in repos, tools like this can actually use them. if everything lives in figma files and google docs that AI can’t access, you’re going to spend more time preparing context than building.

why it matters anyway

even with the limitations, the direction is clear. the bottleneck for startup content has always been production. waiting for someone to make the thing. this shifts it to iteration speed. you build rough, refine fast, and ship in hours instead of weeks.

we’re still early. but i’d rather be figuring this out now than hiring another video editor in six months and wondering why it still takes weeks to get a feature demo out the door.

would love to see what you create!

see you next time!

xx Lisa

useful links:

• remotion docs

• remotion prompting guide

• claude code

hi friends,

i’ve been testing Comet, Perplexity‘s AI-native browser, and it’s starting to feel less like a tool and more like an assistant that lives inside Chrome. (think: an agent inside your browser, not just a sidecar chat)

what makes Comet feel different

• it natively embeds the Perplexity assistant meaning you don’t search in a separate window. you ask, and it acts across open tabs, apps, and workflows, uninterrupted.

• it’s built for action: book a meeting, send an email or navigate forms right from the page you’re on.

• it summarizes, searches and executes: it manages your calendar, briefs you on your inbox, sends messages, and can even nudge things forward.

here’s what i’ve tested with Comet so far

• updating deal notes and creating tasks inside HubSpot on my behalf

• summarizing sales meeting recordings and emailing post-meeting notes

• pulling exact Q3 forecasted numbers from dashboards

• booking an acupuncture session (almost—when accounts are required, it stops and guides you)

• acting as my calendar + email assistant. i asked: “review my last six months of calendar activity and tell me what i should automate.”

• researching prospects for an upcoming event in NYC and adding them to a Google Sheet

• summarizing the last interview on youtube with Dario Amodei

below are real screenshots from my tests, and how Comet performed

sending post-meeting notes

• task: summarize today’s pipeline review and email Andrei, Andrew, Jalena, and Michael.

• result: pulled the meeting transcript from Jimminy, summarized it, and sent the email to the exact recipients.

analyzing my last 6 months of calendar data

• task: analyze my calendar for the last 6 months and tell me what i should automate.

• result: gave me a breakdown of recurring tasks and high-impact areas to automate, from meeting prep to gym sessions.

finding and summarizing our all-hands recording

• task: find in google drive the recording from our last all-hands and run a summary of the Q3 goals to send to the company.

• result: Comet searched my drive, located the right recording, and prepped a summary of our Q3 priorities ready to drop straight into an email to the team.

why this is different: Cristian Tamas from my team said it best:

“Comet runs agents right in your browser—especially helpful when you’re already signed in everywhere. just be careful you don’t accidentally delete something in Hubspot.”

unlike ChatGPT’s agent mode, Comet runs in your browser environment, giving you full control and full context without constant logins or switching tabs.

tl;dr

Comet feels like the start of a new category.

the browser that thinks with you.

it collapses tabs, actions, and ideas into a single flow, moving us toward a true zero-click experience where the assistant handles tasks before you even think to search. in a landscape dominated by ad-driven incumbents, Perplexity’s CEO, Aravind Srinivas, is positioning Comet as built for users, not advertisers which could make the upcoming browser wars very interesting.

there are still limitations, of course. they have some way to go to become truly best-in-class when it comes to authentication, speed, seamless downloading of multiple assets at once, or following through on actions completely without my intervention. but they’re getting there, and the progress is fast.

right now, access is still limited through Perplexity Max with a pro tier rolling out soon, but the direction is clear. the winners will be the browsers that don’t just find information but also send the email, update the CRM, and add the calendar invite, all in the same motion. and the moment you experience that, there’s no going back.

i have a few invites in my account. drop your thoughts and a note if you want one!

xx

Lisa

how to (finally) build institutional knowledge with AI

Lisa Popovici

Co-founder at Siena AI | Helping brands create AI-powered CX | Forbes 30 under 30

June 17, 2025

hi friends,

you know that sinking feeling when someone asks about a decision from last quarter—or even last week—and nobody remembers exactly why it happened? or when new team members spend weeks trying to catch up on historical context?

meet your new, silent superpower: Gemini note-taking in Google Meet.

it doesn’t just save your meeting details but transforms your team’s conversations into a permanent, searchable institutional knowledge base.

why this matters

• never lose context: Gemini automatically transcribes your meetings, captures detailed notes, and suggests clear next steps transforming conversations into searchable, shareable team docs.

• team harmony: everyone invited gets identical notes, zero FOMO, zero misalignment.

• centralized memory: entire team uses the same notetaker and everything is stored in one place. clean, consistent, and searchable.

what you’ll get obsessed about

• Google Workspace integration: Gemini integrates with Calendar, Docs, Drive, Gmail, and Sheets. meeting notes get automatically attached to your calendar events and stored in drive.

• automated workflows: use your Gemini-generated notes to draft follow-up emails or summarize key actions directly in Docs or Gmail. (think of it as your new EA minus the expense report.)

• no extra cost: if you’re already on Google Workspace, it’s free to use.

• the BIGGEST benefit of all: this creates your company’s institutional knowledge. every meeting, every detail is searchable, shareable, and permanently available. and when integrated with ChatGPT, you can literally ask anything anytime and instantly retrieve your team’s historical insights. ultimate self-serve enablement!

how it works

when you add a G-Meet in your calendar, click on the video call options and click the three-dot menu.

select “meeting records” and turn on “take notes with Gemini” (toggle “transcription” too if you like details).

quick tip: make this part of your meeting checklist—non-negotiable. that way, you’ll not forget to turn it on once the meeting starts.

what you’ll get

• clear, actionable notes with summaries, detailed context, and recommended next steps.

• full meeting transcript (which i still toss into chatgpt when i want to dive deeper into a specific topic).

now, i know it’s not as sophisticated as other cool notetakers such as Granola or Circleback (which i love), but it’s just the beginning and long-term, building your company’s institutional knowledge will pay off massively.

seriously, your future self (and team) will thank you.

next week: another quick, actionable AI hack to make work lighter and life easier.

xx Lisa

ok, confession: we all have a love/hate relationship with CRMs.

HubSpot, Salesforce... they’re powerful, essential, and sometimes a bit frustrating to manage day-to-day. you know the drill—endless notes, updating deals, and making sure everything’s perfectly logged.

so here’s the good news: ChatGPT just made everyone’s lives a bit easier. you can now directly integrate HubSpot into ChatGPT to automate (and upgrade) your workflow.

yup, this is as cool as it sounds.

here’s what this means for you (and me):

• automatic deal summaries: imagine having a quick, clear summary of every sales call, note, or email conversation ready to go. no more manual recaps (and no more excuses for not logging deal updates, friends).

• email generation magic: create personalized follow-up emails directly from your CRM. Hubspot + ChatGPT instantly drafts messages tailored to your prospect’s context and conversation history—saving hours.

• sales and marketing alignment: keep everyone in sync by generating weekly pipeline updates, deal insights, and even internal briefings automatically.

• supercharged customer insights: turn lengthy interactions into concise insights, so your teams always have context at their fingertips.

• deep sales efficiency analysis: instantly analyze deal velocity, win-loss patterns, and team performance to pinpoint exactly where you can improve your sales efficiency.

quick setup guide (promise it’s easy):

open ChatGPT (Pro/Enterprise) and click on the tools menu at the bottom. select “run deep research.” this enables the new data analysis mode inside.

click on sources. you’ll see some default sources (like web, PDFs, etc.) and a “connect more” option. tap that, and find HubSpot in the list of available connectors.

authorize the connection. a HubSpot login window will pop up. log in and choose the HubSpot account or portal you want to link (if you have multiple). ChatGPT will confirm once the connection is made.

start asking away! back in your ChatGPT chat, you should now see a toggle for HubSpot under sources. make sure it’s switched on. from here, anything you ask while in deep research mode will use your HubSpot CRM data as context.

pro tip:

try my favorite combo: o3 model + research mode + HubSpot + Linear 🤸♀️

see you next week—there’s a lot of cool new integrations available on ChatGPT now. i’ll write about notetaking with Gemini, Google Workspace, Linear and more.

xx Lisa

today i’ll talk about promotions, but not the slow kind. not the “i’ve been here for three years and it’s finally my turn” kind.

i mean the fast ones. the kind where someone’s scope expands so clearly that the title change almost feels like paperwork catching up to reality.

over the last year, i’ve watched a people get promoted not once, not twice, but multiple times. and it’s been consistent across functions: sales, CX, ops. different roles, same pattern. and what’s interesting is it wasn’t because they were working insane hours, or because they were louder, or because they were perfect. the common thread was simpler: they turned AI into leverage for the organization, not just for themselves.

most people are already using AI quietly. they’re drafting faster, thinking clearer, prepping better, shipping more. all of that is real value. but it’s also mostly invisible. and invisible work rarely gets rewarded the way it should.

the people who advance fastest tend to do something slightly different. they take what they’ve learned privately and turn it into something other people can use. a workflow. a template. an agent. a checklist. a feedback loop. anything that doesn’t just make them faster, but makes the team faster.

i recently listened to an episode of how i ai that made this click even harder for me. it featured Zapier’s CEO Executive Assistant, Cortney Hickey, and her story honestly stopped me. she’s been promoted three times. and it wasn’t because she “uses AI.” it was because she used AI to redesign how work happens around her.

one example: she built a weekly meeting prep agent. every friday morning it runs automatically, scans her exec’s calendar for the upcoming week, identifies which meetings actually need prep, then does the research she would normally do manually. it looks up external participants, checks the CRM for relationship history, searches email and slack for context, and then outputs two things: a clean prep digest in slack and actual prep tasks in her to do list, timed to show up before each meeting.

she uses Zapier (obviously duh) for her agents and workflows. you can build these in Gumloop or similar tools as well.

but the point isn’t the tools. the point is what she did with her knowledge of the job. she took a messy, repetitive, high-context process that lived in her head and turned it into a repeatable system.

then she went further. she built a meeting coach that sends post-meeting feedback tied to company values and operating principles. not in a weird, performative way, but as a consistent reinforcement mechanism. the kind of thing companies say they want, but rarely execute continuously because it’s too manual and socially awkward. because it’s automated and “neutral,” people actually receive it better, and over time it normalizes feedback in a healthier way.

and then she built an internal doc review GPT that helps teams stress test strategy docs before they ever reach execs. so people come in sharper, with clearer tradeoffs and cleaner recommendations. sometimes it even makes the meeting unnecessary because the doc becomes good enough to decide asynchronously. that part alone is a massive unlock in any org.

we’re seeing this play out very clearly in CX too. over the past year, several leaders using Siena didn’t just adopt AI quietly, they turned it into shared infrastructure for their teams.

at Momentous, CX was an early adopter of AI, and that leadership created a ripple effect where other departments started following the same playbook.

at Comfrt, deeper AI usage translated into broader ownership across customer experience and operations.

at True Classic, AI became a lever to scale decision-making, automation, and team impact beyond day-to-day execution.

at Brooklinen, sustained investment in CX, AI and automation led to expanded scope and more strategic responsibility.

in every case, the progression wasn’t about replacing work but about mastering AI well enough to redesign workflows, raise the bar for the team, and operate at a different level. when AI moves from a personal shortcut to shared infrastructure, titles change because the job itself changes.

these leaders used AI to make their teams better, more consistent, and more scalable. that’s the difference between private AI usage and coordinated adoption done right: when institutions reward it, AI stops being a quiet advantage and starts becoming a career accelerator.

i keep thinking about the same underlying idea: private AI makes you faster. visible AI makes you promotable.

when you use AI only for yourself, you get efficiency. when you turn it into shared infrastructure, you get leverage. and that’s what leaders actually reward, because it compounds across teams.

also, AI doesn’t implement itself. the real skill is being able to explain the system behind your work. why you do things the way you do. where context lives. what “good” looks like. what can go wrong. what needs judgment. that’s why the EA example is so powerful. she didn’t just automate tasks. she scaled judgment.

and that’s the part that makes you hard to replace, not easy to replace.

i know some people hesitate here. there’s a quiet fear that if you automate too much or share too much, you make yourself replaceable. but what tends to happen in practice is the opposite. the person who hoards AI gains becomes an individual contributor with better output. the person who shares them becomes the reference point for how work gets done. leaders don’t promote the fastest typist. they promote the person who makes everyone else faster.

if you want to apply this without overthinking it, here’s a simple way to start this week. pick one repeated workflow that matters to your team. something you do every week that takes time and brainpower. write down how you currently do it step by step, like you’re explaining it to a smart intern. then build a basic AI version of it. it does not need to be perfect. share it with one person. iterate once based on what breaks. then make it visible.

you don’t need permission. you need clarity and follow through.

my biggest takeaway from watching this pattern repeat is that AI doesn’t replace roles. it replaces bottlenecks. and the people who identify bottlenecks and remove them tend to move up very quickly.

i’m going to keep sharing examples like this because they’re more useful than tool lists or hype. and if you’re building something like this inside your own company, i’d genuinely love to hear about it.

i know you’re already crushing your 2026,
xx Lisa

hey friends,

today’s newsletter is short and sweet.

i watch a ton of long-form content. podcasts, founder interviews, strategy conversations. Lenny Rachitsky‘s podcast especially, when i want to see how operators actually think and action.

for a long time, i’d watch, take notes manually, pause and rewind, then either share with the team or apply to my work at Siena AI. it worked, but it was slow. i was spending time capturing information instead of synthesizing it.

what changed is how i’m using Perplexity‘s comet assistant.

comet sits in the browser while i watch. instead of extracting signal in real time, it does the first pass for me.

example: Elena Verna‘s recent appearance on Lenny’s podcast. (if you don’t know Elena, she’s one of the sharpest growth leaders out there. she’s helping Lovable scale fast and she’s clear about how classic growth playbooks are breaking in AI. huge shoutout to her and Lenny for consistently surfacing operators doing the work.)

while watching, instead of taking notes, i asked comet:

• what are Elena’s key takeaways?

• what is she spending time on now vs before?

• what changed in how she thinks about growth in AI?

comet pulled the signal immediately. things like: she’s spending way more time on rapid product innovation, distribution, and storytelling. way less on deep funnel optimization and incremental experiments. growth is now driven by speed, wow moments, and product love. community and narrative matter more than perfect attribution.

then i go one layer deeper. i ask: how does this apply to an AI company like Siena? what does this mean for how we ship and position? what should i stop over-optimizing?

suddenly the content isn’t just interesting. it’s operational.

instead of passively consuming a 90-minute podcast, i turn it into decisions, conversations, and experiments.

the biggest unlock is speed. i can watch more, learn more, stay focused on building. i’m not blocked by note-taking.

it also changes how i share learnings internally. instead of forwarding a long video saying “this is worth watching,” i share a few sharp insights that actually matter, plus the moments worth revisiting.

AI does the first draft of understanding. i do the judgment, context, and decisions.

if you want to try: watch something you care about, ask comet for main takeaways, ask one follow-up tied to your role, apply one insight immediately.

tools like this aren’t flashy. but they compound quietly. the tighter your learning loop gets, the faster you move.

happy holidays! 🤗🎄

xx Lisa

one thing you quickly realize as a fast-growing company: moving at high speed is messy.

the truth is:

• internal knowledge constantly evolves

• customer use cases keep growing in complexity

• team members often find themselves digging through countless documents or waiting on responses from other departments

this friction can easily slow your team down and, worse, frustrate your customers and partners.

at Siena AI, we’ve been there. and since we’re all about empathic AI, we thought—why not make Siena our own internal agent?

so, we did just that.

now, when our team, customers, or partners have questions, they don’t wait for answers—they just ask Siena in slack.

here’s what this looks like in practice (real examples from our Slack channels):

• partner support enablement. we recently onboarded our new partner, TalentPop. instead of asking our busy team, they drop questions directly to Siena in Slack, who instantly responds. kudos to Michael Potters for building this!

• customer enablement. customers like Katie Mitchell and Sierra DiMaggio feel understood while having a good laugh lol. shoutout to Chad Horenfeldt and Svetlana Guzovskaia at Siena for building this!

• team self-enablement. our team does so much hands on work for our customers. Siena keeps pace by getting them instant answers.

how does Siena stay smart?

great enablement relies on accurate, updated, and easily accessible information.

• Zendesk knowledge center: this is our Siena University aka our central hub for customer-facing docs, FAQs, and external resources.

• Cursor: our internal IDE (Integrated Development Environment) and knowledge hub. our team taps directly into Cursor to access our latest internal docs, project notes, and technical insights, ensuring everyone stays aligned and productive.

• Linear + Anthropic (Claude): we’ve integrated our product management tool (Linear) with Claude, ensuring the team always has the freshest product updates, roadmap items, and feature insights straight from the source.

this setup means our teams can instantly self-enable, staying informed, autonomous, and able to move at AI-era speed.

why enablement via AI matters (a lot):

• speed: enable your teams to move autonomously, at their own pace, without bottlenecks.

• accuracy: your knowledge is centralized and consistently updated—no misinformation.

• satisfaction: instant, reliable support drives happy customers (and partners).

• empowerment: less dependence, more confidence.

as you grow, complexity grows too—enablement must scale. today, being able to quickly self-enable isn’t optional anymore.

how to get started:

• train your AI agent on all internal docs and support resources. it should be your single source of truth.

• encourage direct interaction: let everyone—customers, partners, teams—chat directly with the agent.

• continuous learning: regularly update your knowledge documentation so the agent is always up to date.

• set clear expectations: your agent won’t be perfect overnight. clarity, testing, and ongoing iteration are key.

now go build yours. once you have it, you’ll never go back.

x Lisa

PS: shoutout to my wonderful friends Grace Clarke and Éva Goicochea for hosting the 6th Founders Weekend retreat this past weekend in Sancerre, France. in a world moving faster than ever—especially in AI and our industry—there’s nothing more important than having a tribe of like-minded, inspiring people who genuinely support, cheer, and listen to each other. grateful to have found mine.

this question keeps execs up at night.

i get it—the AI wave is happening faster than anyone anticipated. falling behind isn’t an option, and making a hasty hire can burn resources without clear results.

the good news? you probably already have more than you realize internally.

here’s exactly how we did it at Siena AI (and what i recommend all my friends do):

first things first—embed AI into your culture

• make AI fluency a non-negotiable company value. update your core values if needed to reflect curiosity, creativity, and adaptability, making AI not just an initiative, but who YOU are.

• get leadership bought in first. sell the vision of becoming an AI-first organization—explain clearly how it elevates the business and everyone’s careers. if leaders aren’t excited, no one else will be.

now, get the whole team involved:

• set up dedicated slack channels (we use #ai-at-siena, #ai-watercooler, #ai-cursor, #knowledge-search). these quickly became our most active channels.

• run monthly AI workshops. if you don’t have all the answers, bring in friends, partners, or experts to lead these sessions. give your team a continuous learning path.

• challenge manual workflows. whenever you see repetitive tasks, ask:“why aren’t we automating this with AI yet? what’s the real barrier?”. encourage the team to question manual processes instead of just accepting them.

• challenge headcount requests: anytime someone requests new headcount, ask first: “what would this project or role look like if AI handled 90% of the tasks? why can’t this be done using AI? if it can’t be done with AI, can we use contractors or agencies?” this approach has sparked some of our best projects and discussions.

• leaders should regularly review team outputs. make sure everyone is leveraging AI. if work is consistently being done without AI, that’s a red flag—it means there’s untapped potential, wasted time, and unnecessary friction.

i love how clearly Tobias Lütke put this to his team:

“using AI effectively is now a fundamental expectation of everyone at Shopify. frankly, i don’t think it’s feasible to opt out of learning the skill of applying AI in your craft; you are welcome to try, but i want to be honest i cannot see this working out today, and definitely not tomorrow. stagnation is almost certain, and stagnation is slow-motion failure. if you’re not climbing, you’re sliding.”

exactly this. AI fluency isn’t a “nice-to-have”—it’s quickly becoming table stakes.

here’s why training internally makes sense first:

• your current team deeply understands your customers, your business, your brand—they already have context. AI tools amplify existing talent, not replace deep expertise.

• economist Tyler Cowen predicted this in “average is over”: AI divides the workforce—one group leveraging tech to achieve exponential results, and another struggling to keep up. training your team puts everyone in the first group.

as my co-founder, Andrei Negrau , said perfectly in his recent substack:

“the divide isn’t between those who use AI and those who don’t. it’s between those who understand their craft deeply enough to direct AI with precision—and those who don’t.”

signs you might be ready for an external AI leader:

• complexity increases exponentially—multiple teams using AI in different ways without cohesion.

• scaling is bottlenecked due to lack of strategic alignment.

• internal champions need someone focused full-time on AI vision and execution.

until then, focus internally. invest in your people. the best hires come later, naturally, as you outgrow current capabilities.

how to upskill your team starting today:

• lead by example. share your own AI experiments openly to create a culture of psychological safety around new tech.

• provide premium subscriptions (claude pro, chatgpt pro, perplexity pro, cursor, lovable etc.). think of these as essential tools—not luxuries.

• encourage hands-on experimentation. give permission to fail, learn, and iterate.

my top 3 favorite learning resources

• Greg Isenberg

• Lenny’s podcast

• How I AI podcast

the truth is: AI is fundamentally changing every role and hiring process, but it doesn’t replace deep expertise. so why not adapt and thrive!

the real advantage isn’t having the latest AI exec on your team (honestly these roles are extremely hard to even find). it’s having people who deeply understand your business and how to wield AI to deliver results at scale.

share below your team’s biggest AI challenge—i’ll give you more practical tips.

bisous,

Lisa

bonjour mes amis,

last week in our sales monthly, we identified a pattern: due to being last month of the year and 1 week before xmas, some deals are taking longer to cross the finish line. prospects love the product, understand the value, but push the adoption to next quarter. we needed a way to make the cost of waiting tangible.

so i built a cost of inaction calculator. not by hiring someone or opening a ticket with engineering. i built it myself in a conversation with claude, using artifacts.

what artifacts actually are

artifacts let you build real, working tools directly in a chat interface. you describe what you want, claude builds it, you see it live, iterate on it, and download the final product as a standalone file. no coding required.

for this calculator, i told claude what inputs we needed (ticket volume, handle time, agent cost), what we wanted to show (daily/monthly/annual cost of waiting), and how it should look (clean, black and white, professional). 30 minutes later, i had a fully functional web calculator our AEs can send to prospects.

why this matters for sales leaders

i’ve been in sales long enough to know that the tools you need don’t always exist. your team identifies a gap, you submit a request, it goes into a backlog, maybe you get something in six months. by then, the problem has evolved or the urgency is gone.

with artifacts, you can build what you need when you need it. sales enablement tools. ROI calculators. pitch materials. comparison charts. whatever closes deals faster.

the actual use case

so if you are experiencing longer deal cycles to “let’s revisit this next quarter.” in Q4, it’s not because prospects didn’t see value, it’s because not everyone is realizing the urgency. you need to quantify what waiting costs them.

the calculator does exactly that. prospect inputs their CX metrics (tickets per month, handle time, team size, hourly cost). it shows them in real dollars what they’re losing every single day they wait to automate. $xxzx per day. $yyyyy per month. $zzzzzzz per year.

suddenly “let’s wait until Q1” becomes “we’re bleeding $100k over the next 90 days if we wait.”

how we’re using it

• demo follow-ups when timing becomes an objection

• nurture sequences for stalled pipeline

• live on discovery calls to build urgency

• shared directly with economic buyers who care about cost

it’s not a replacement for good sales process. but it’s a forcing function that makes prospects confront the cost of inaction in concrete terms.

why sales leaders should care about AI tools

i’m not technical. i don’t code. but i can build what my team needs to close deals faster. that’s the unlock here.

most sales leaders are dependent on other teams to build enablement tools. with AI, you’re not anymore. you can prototype, test, iterate, and deploy tools yourself in the time it used to take to write the spec document.

we’re moving into a world where the constraint isn’t “can we build this?” but “do we know what we need?” if you can articulate the problem clearly, you can build the solution.

how we’ll know if it actually works

building the tool was the easy part. it took 20 minutes. the hard part is knowing if it moves deals or just creates busy work.

here’s how we’re tracking success:

week 1-2: adoption are AEs actually using it? if they’re not reaching for it in real conversations, the tool doesn’t matter. we’re tracking how many times it gets shared in deals and collecting qualitative feedback from the team.

30-60 days: deal velocity we added a Hubspot field: “cost calculator shared: yes/no.” after 60 days, we’ll compare conversion rates and average days to close for deals where the calculator was used vs deals where it wasn’t. if deals move faster when prospects see the numbers, we have our answer.

ongoing: pipeline health we’re watching for timing objections. are we still hearing “let’s move our meeting to January” as much? are deals stalling on implementation timelines? if the calculator works, those objections should decrease.

the real test if AEs stop using it after a few weeks, it doesn’t work. tools that actually help close deals get adopted naturally. no amount of enablement sessions or Slack reminders will fix a tool that doesn’t resonate with prospects.

if they keep using it and deals accelerate, we’ll know we built something that matters.

that’s the thing about building fast with AI - you can test, learn, and iterate without massive investment. if this calculator works, great. if it doesn’t, we’ll build something else. the constraint isn’t resources anymore. it’s speed of learning baby.

that’s how GTM should work today.

à la prochaine!

xx Lisa