Because the product is still in beta, some button names and screens may change.

1. Check that you have access

At launch, Grok Bot access is being offered through selected paid plans, reportedly including SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium.

Go to the official Grok Bot download page. Sign in with the account connected to your eligible subscription. If you do not have access yet, the page may ask you to upgrade or join a waitlist.

Grok Bot is currently available on desktop and iOS, with availability expected to expand as the beta develops.

2. Create your first bot

Open Grok Bot and select Create a New Bot or the + button.

The app will ask what you want the bot to focus on. Start with one clear, repeatable job instead of asking a single bot to manage everything.

For example:

You are my email assistant. Review new partnership inquiries, summarize the important details, and draft concise replies in my tone. Never send an email without asking me first.

Grok Bot may ask follow-up questions about your workflow, preferred output, and approval requirements. Answer these as specifically as possible.

3. Give the bot clear instructions

A useful bot setup should define five things:

• Role: What kind of assistant is it?

• Responsibilities: Which tasks should it handle?

• Sources: Which apps, websites, or files may it use?

• Output: How should it present its work?

• Boundaries: What must it never do without approval?

Here is a reusable setup template:

You are my [role]. Your main responsibility is to [primary job].

You may use [approved apps, files, and websites].

When completing a task, provide [desired output or format].

Always ask for approval before [sending messages, publishing, purchasing, deleting, or changing important information].

If information is missing or uncertain, ask me instead of guessing.

4. Connect the apps it needs

Open the bot’s Plugins, Apps, or Connections area. Select the service you want to connect, such as Gmail, and complete the sign-in and permission process.

Only grant access that the bot genuinely needs. For example, an email-drafting bot may need permission to read messages and create drafts, but it should not need permission to send emails automatically.

If a direct integration is unavailable, the bot may be able to access the service through its cloud browser. Sign in inside that browser when prompted. Keep in mind that browser-based connections may be slower or less reliable than direct app integrations.

5. Add safety rules

Before letting the bot work independently, set clear approval boundaries. A good beginner setup is:

• It may research, summarize, organize, and draft independently.

• It must ask before sending messages or publishing content.

• It must ask before spending money or making purchases.

• It must ask before deleting or permanently modifying anything.

• It must stop and report back if it encounters a login problem, CAPTCHA, or unexpected screen.

• It must never share passwords, private files, or sensitive information.

Avoid giving your first bot access to financial accounts, confidential records, or other high-risk systems.

6. Run a small test

Give the bot a short, low-risk assignment before creating recurring work.

For example:

Review the five newest emails in my inbox. Identify anything urgent and draft replies where appropriate. Do not send or delete anything.

Check whether the bot:

• Understood its role

• Used the correct account

• Followed your approval rules

• Produced useful work

• Stopped when it encountered uncertainty

If something went wrong, correct it directly: “From now on, do not draft responses to newsletters,” or “Keep summaries under five bullet points.”

7. Schedule recurring work

Once the bot completes the task reliably, ask it to repeat the workflow on a schedule.

For example:

Every weekday at 9:00 a.m., review new partnership emails, prepare a short priority summary, and draft responses. Leave every response as a draft for my approval.

Start with one scheduled task and monitor it closely. Cloud-computer work can consume usage quickly—especially if a bot gets stuck, reloads pages repeatedly, or encounters login problems.

8. Build a small team of specialized bots

Instead of creating one bot that handles your entire business, create several focused bots:

• Inbox Bot: Sorts messages and prepares drafts

• Research Bot: Produces a daily industry brief

• Content Bot: Turns research into post ideas

• Operations Bot: Reviews tasks and upcoming deadlines

• Quality-Control Bot: Checks another bot’s work before you approve it

Single-purpose bots are easier to test, improve, and trust. Early users also report that only a limited number of bots may work simultaneously, so begin with two or three instead of building a large team immediately.

Troubleshooting

I cannot access Grok Bot:
Confirm that you are signed in with the account attached to an eligible plan. The product is in early beta, so access may still be limited.

A connected account keeps signing out:
Reconnect it through the app’s integration or plugin area. Browser-based sessions may occasionally need to be restored.

The bot is using too many credits:
Stop the task, shorten its instructions, reduce the number of websites it checks, and add a rule telling it to stop after a fixed number of attempts.

The bot is taking actions I did not expect:
Remove unnecessary permissions and strengthen its approval rules. Test again with a small task before restoring scheduled work.

The bot gets stuck on a website:
Ask it to stop and report the exact problem. CAPTCHAs, login checks, and changing interfaces may require you to take over manually.

Best practice

Treat Grok Bot as a capable new teammate—not an unsupervised employee. Give it one clear responsibility, limited access, and firm approval rules. Review its work closely at first, then increase its independence only after it performs the workflow reliably.

Grok Bot is still an early-beta product, and initial user reports mention browser-session issues, limited concurrency, and potentially high usage during computer-control tasks. Start small and avoid making it the only system responsible for important business operations.

You’re not unemployed. You’re not even doing badly. On paper it looks fine — the job, the title, the thing you say when someone asks what you do. And you’re lying in the dark running the same loop: is this it, did I pick wrong, is there a version of this where I’m actually good at something.

Then you open LinkedIn, which is the equivalent of treating insomnia with espresso.

I spent a long stretch of my twenties in that loop. And the thing I eventually figured out is that I was searching in the wrong category. I kept googling roles. Salary bands. “Careers for people who like X.” I was treating it as an information problem — like the right job title existed somewhere and I just hadn’t found the search term yet.

It wasn’t an information problem.

You’re not confused about which job to take. You’re confused about who you’re allowed to become. Those are different problems. Only one of them has a job board.

I want to walk through five essays that got me out of that loop. Not as a reading list — reading lists are procrastination with a productivity costume on. I want to show you the single thing they all quietly agree on, which none of them says outright, and which took me embarrassingly long to see.

1. Most of your ambition was installed by someone else

Tim Urban’s How to Pick a Career (That Actually Fits You) is long enough that you’ll be tempted to skim it. Don’t.

His core move is to separate what he calls your “authentic voice” from everything layered on top of it. The expectations you absorbed from your parents. From your degree. From the specific five people in your group chat. From whatever version of “impressive” your industry runs on.

The part that landed for me: you don’t experience any of that as external pressure. You experience it as wanting things. It feels like your own taste. It feels like intuition. That’s exactly why it’s so hard to audit — you’re running software you never wrote and calling it a personality.

Here’s a test that costs you nothing. Think of the career outcome you most want. Now ask who, specifically, you’re imagining telling about it. If a face shows up fast and clear, that goal has a co-author.

That’s not a reason to throw the goal out. Plenty of inherited goals are good ones. But you should know which ones you chose and which ones were handed to you, because the handed-down ones are the ones you’ll grind at for a decade and feel nothing when you finally hit.

2. Clarity comes after action, not before it

This is the one that actually changed my behavior, so I’ll be direct about it.

Herminia Ibarra studies career transitions. Her finding, across years of research on people who successfully changed direction, contradicts basically everything we’re taught: people did not figure out what they wanted and then act. They acted, and the acting is what produced the clarity.

We’ve got the sequence backwards. The standard model is: reflect deeply → arrive at your true calling → execute. Take the personality test. Journal about it. Wait for the feeling of knowing.

That model has a fatal flaw, which is that you cannot introspect your way to information you don’t have yet. You have no idea whether you’d like running a small team, or writing for a living, or working in a field you’ve only read about. Not because you haven’t thought about it hard enough. Because you’ve never done it, and doing it is the only source of that data.

Ibarra’s alternative is what she calls possible selves — running several small, low-cost experiments in parallel rather than making one big irreversible decision.

Concretely, in your life, that looks like:

• A side project you ship publicly, small enough to finish in a weekend

• A freelance engagement in the field you’re curious about, even a tiny one

• Reaching out to someone doing the thing and asking to help with something real

• Volunteering for the internal project nobody wants, in the department you’re curious about

None of these are commitments. They’re probes. You’re not trying to find the right answer, you’re trying to generate real reactions in yourself that you can’t fake or predict from your couch.

And the signal you’re watching for isn’t “did I love it.” It’s which one did I not want to put down. Which one did I keep thinking about in the shower. Which one made time behave weirdly.

That’s data. Journaling is not data.

3. Competent is not the same as strong

Peter Drucker’s Managing Oneself is from 1999 and it’s shorter than most of what you’ll read this week.

Drucker’s question isn’t “what are you good at.” It’s sharper: where do you produce results that other people can’t easily produce?

Most of us conflate two very different things. There’s competent — you can do it, it works, nobody complains. And there’s strong — you do it noticeably better than the people around you, with less effort than it seems to cost them.

Competent is what gets you hired. It’s also a trap, because being competent at something is remarkably pleasant. You get praise. Nothing goes wrong. Years pass.

Strong is the only place where compounding happens. Ten years of building on a genuine strength produces something categorically different than ten years of being adequate at four things.

The hard part is that you’re the worst-positioned person to identify your own strengths, because your strengths feel like nothing to you. They’re the things you assume everyone can do. Drucker’s answer is feedback analysis: write down what you expect to happen before you take on something significant, then compare it against what actually happened, months later. Do that for a couple of years and your real pattern becomes undeniable.

The cheap version, if you don’t want to wait two years: ask five people who’ve worked with you what they’d come to you for specifically, and what they’d never come to you for. The second answer is the useful one.

4. Excitement plus aptitude, not passion

Paul Graham’s How to Do Great Work is for a specific reader — the person who is technically doing fine and is secretly, corrosively bored.

His argument is that great work sits at the intersection of what you’re genuinely excited by and what you have unusual aptitude for. Note that “passion” is not the word. Passion is a feeling you’re supposed to already have, which is precisely the thing you don’t have and are anxious about. Excitement is smaller and more useful — it’s curiosity that keeps showing up unprompted.

The second half of his argument is the part people skip: you find that intersection by working on things. Interest is not something you discover by consulting yourself. It grows on contact with actual problems. You get mildly curious, you poke at it, and the curiosity either dies or it deepens into something with real gravity. There’s no way to run that simulation in advance.

Which is the same thing Ibarra found, arrived at from a completely different direction. Two people — one studying career transitions empirically, one reflecting on how ambitious people produce great work — landing on the identical mechanism.

Graham also gives you the honest cost, which most career advice won’t: this takes years. Not a quarter. Years of consistent effort in a direction, which is why picking a direction you’d tolerate being wrong about for two years matters more than picking the theoretically optimal one.

5. What are you actually building

Clayton Christensen’s How Will You Measure Your Life? zooms all the way out, and it’s the one that stays with me.

He noticed something about his Harvard Business School classmates. Extraordinarily capable people. And a striking number of them ended up in lives they had never actually chosen — divorced, estranged from their kids, deeply successful and quietly miserable.

His diagnosis is that we allocate our resources — time, attention, energy — toward whatever gives the fastest, loudest evidence of progress. Work does that. Work gives you a promotion, a launch, a number that went up, a message that says you did well. Relationships, health, and the long slow work of becoming someone don’t send those signals. They return on a ten-year timeline and pay you nothing in the meantime.

So the marginal hour goes to work. Every time. Not through one bad decision but through ten thousand entirely reasonable ones.

The reason this belongs in a career essay: the career question was never really “which job.” It was always “what life am I constructing, and is this job a load-bearing piece of it or is it quietly eating the rest.” You can optimize a career perfectly and still be answering the wrong question for thirty years.

What they actually agree on

Read those five and the shared claim underneath them is uncomfortable:

There is no answer waiting to be found. There is only a process, and it runs on action rather than thought.

Urban says your stated preferences are contaminated, so don’t trust them at face value. Ibarra says reflection doesn’t resolve the contamination — experiments do. Drucker says your strengths are invisible to you until evidence accumulates. Graham says interest deepens through contact with work, not contemplation of it. Christensen says even a well-run career drifts unless you check it against what you’re actually building.

Not one of them tells you what to do. Every one of them tells you to go generate data on yourself.

That’s frustrating if you came looking for an answer. It’s freeing if you’ve been treating your indecision as a personal failure — because it reframes the whole thing. You didn’t fail at choosing. You were running a method that doesn’t work, that everybody handed you, that nobody checked.

If you want to actually do something with this

The order matters, so here it is as a sequence:

Urban → Ibarra → Drucker → Graham → Christensen.

Understand your wiring. Run experiments instead of making decisions. Find where you’re strong, not just capable. Go deep on it. Then check it against the life.

And a starting move, because reading five essays and changing nothing is the most likely outcome here:

Pick one possible self. The most curious, least justifiable one. Give it a deadline four weeks out and a scope small enough to actually finish. Ship it publicly, or at least to five people. Then write down — before you start — what you expect to feel, so you can check it afterward against what you actually felt.

That’s one data point. You need maybe six.

You were handed a major at nineteen and told to figure it out. Nobody taught you the method. Fifteen years of “follow your passion” is not a method, it’s a poster.

Go get the data.

I’ll see you next week with a new essay. :)

Week of June 22–26, 2026 · from @creativelyange

Hey friends 👋

Big week. Figma basically reinvented itself, AI tools got more expensive (yeah, you read that right), and the design layoff conversation got louder. Here’s everything I covered this week — skim it in two minutes.

🎨 Figma Config 2026 went kind of insane

This was the headline of the week. A few things dropped at once:

• Motion — you can now animate inside Figma. No After Effects, no handoff. Animate a component once and the motion follows it across every screen.

• AI shader fills — type “frosted glass” or “liquid metal” and it builds a real, editable effect on your canvas. No code.

• Plugins from a sentence — describe the tool you wish existed and Figma builds it for you.

• Aleph 2.0 (Runway’s video model) now lives in Figma — you can direct AI video frame by frame, swap a character, change the background, no reshoot.

My take: the gap between “idea” and “moving thing” just collapsed. Also — fun fact — the AI powering all this isn’t even Figma’s own model. They’re renting brains from OpenAI, Anthropic, and Google and keeping the canvas. 👀

💸 AI is officially getting more expensive

For three years these tools got cheaper every few months. That just reversed. The newest ChatGPT model is double the last one, Google’s fast model jumped ~3x, and Claude’s new tier launched at double the previous price.

Why? OpenAI reportedly spent about $1.35 for every $1 it made last year. That math was never going to last.

My take: the “use everything, it’s basically free” era is ending — but the cheap option didn’t disappear, it moved (hello, open Chinese models). You don’t have to spend more. You just have to get pickier: expensive model for the hard stuff, cheap one for everything else.

🧠 Anthropic is quietly raiding Google’s AI team

Google lost 4 of its top AI people in 6 days — including a Nobel laureate and Google’s own AI coding lead. Most of them landed at Anthropic (the company behind Claude).

My take: watch where the people go — the products follow. If you’ve been wondering why Claude keeps showing up in design and building workflows, this is the why.

⚠️ The AI layoffs are hitting designers now

This one’s heavier but worth talking about. Google Cloud cut ~100 UX/design roles, and companies like Coinbase are testing “one-person teams” — one human doing the design, engineering, and product. Roughly 1,100+ tech jobs a day are being cut in 2026, with AI named as the reason.

But here’s the catch: a lot of economists think companies are using AI as a cover story for cuts they already wanted (it even has a name now — “AI washing”).

My take: don’t panic, expand. The job isn’t disappearing — it’s getting wider. The people surviving aren’t the best pixel-pushers; they’re the ones who can also ship and direct AI instead of competing with it. That’s a learnable skill.

🖼️ Two tools worth trying this week

• Adobe Firefly Creative Agent (public beta) — you can now basically talk to Photoshop. “Remove the background, brighten her face, export for Instagram” → done. Works across Premiere, Illustrator, InDesign too.

• Canva AI 2.0 — describe a goal and get back a fully editable layered design, not a flat image. Easily the most beginner-friendly of the bunch.

My take: the skill is no longer knowing where the buttons are. It’s knowing what good looks like and asking for it clearly. That part doesn’t get automated.

🎯 The one thing to remember

Tools are getting more powerful and more expensive at the same time. The winners aren’t the people with the biggest tool budget — they’re the ones who know exactly which tool to reach for and how to direct it.

Learn the tools before they’re a requirement. That’s the whole move.

See you next week 💛 Angel · [@creativelyange on TikTok & Instagram]

Enjoyed this? Forward it to someone who’s trying to keep up with AI without the jargon.

Then i said , turn this into an image of me with these suggestions and also draw out what the differences are /what i need.

Feel free to try this out and let me know how it goes for you !

I’m hosting a free live workshop this Tuesday called From Vibe to Product.

It’s for creative, non-technical people who have ideas for apps, tools, workflows, or AI products but don’t know how to actually turn them into something buildable.

Over the past few months, I’ve noticed a pattern:

A lot of people are excited about AI. They’re saving prompts, trying new tools, watching tutorials, and seeing other people launch products.

But when it comes time to build something of their own, they get stuck.

Not because they aren’t smart.
Not because they don’t have ideas.
And not because they need to become software engineers.

They get stuck because the idea is still too fuzzy.

They don’t know:

• who they’re building for

• what problem they’re solving

• what the smallest useful version should be

• how to explain the idea clearly

• what to ask AI to help them build first

That’s what this workshop is about.

In From Vibe to Product, I’ll walk you through how to take one messy idea and turn it into a clear product direction.

By the end, you’ll have:

• a specific audience

• a real problem worth solving

• a simple product concept

• an MVP scope

• your first vibecoding prompt

This is not a traditional coding workshop.

It’s for creators, designers, students, freelancers, coaches, aspiring founders, and anyone who wants to use AI to build products, not just generate content.

I’ll also share more about my upcoming 5-week live cohort, Vibecode Your First AI Product, where we’ll go deeper and actually build, validate, and launch a first AI-powered product together.

The free workshop is the best place to start if you’re curious.

Join the free workshop here:
Luma Link

See you there,
Angel

People are paying $150–$500 for professional color analyses and personal styling sessions. Meanwhile, ChatGPT just quietly became the most accessible personal stylist on the planet — and nobody sent you the memo.

OpenAI dropped GPT Image 2 in April 2026, and it’s not your old DALL-E. This model actually reasons before it generates. It reads your face, your skin tone, your undertones, and produces visual comparisons that look like they came from a professional stylist’s mood board.

The beauty community on TikTok figured this out fast. “AI glow-up” prompts went viral overnight. But most of the prompts floating around are mid — they give you generic results because they’re too vague.

I spent time testing and refining 4 specific prompts that consistently produce the best results. These cover your full personal style stack: color season, ideal hairstyle, makeup placement, and fashion direction.

Here’s the playbook.

Before You Start: The Selfie That Makes or Breaks Your Results

Your output is only as good as your input. GPT Image 2 is incredibly capable, but it needs a clean photo to work with.

The ideal selfie for these prompts:

• Natural light only — stand near a window or go outside. No ring lights, no overhead fluorescents

• No makeup — bare face gives the AI the truest read on your skin tone and undertones

• Hair pulled back if possible — so it can see your full face shape

• Plain background — white or neutral wall. No busy patterns

• Front-facing, shoulders up — think passport photo framing

• No filters — this should go without saying, but the AI needs your real colors

Snap the photo, upload it to ChatGPT, and then paste in the prompts below one at a time.

Prompt 1: Personal Color Analysis

This is the big one — the prompt that tells you which colors make you look alive and which ones wash you out. It uses the Dressing Your Truth (DYT) system, which breaks coloring into four types based on your skin’s undertone, contrast level, and overall energy.

The Prompt:

Using my uploaded photo, act as a professional color analyst and create a personalized color analysis image.

Use the Dressing Your Truth system. Show me draped in the Type 1, Type 2, Type 3, and Type 4 palettes side by side. Keep my hair color exactly as-is (do NOT change my hair). Apply natural, everyday makeup appropriate for each type.

Layer and drape all the colors of each type across my shoulders so I can visually compare how each palette interacts with my skin.

Use the correct DYT color palettes:

• Type 1: light, bright, warm (coral, peach, turquoise, light yellow)

• Type 2: soft, muted, cool (dusty rose, lavender, sage, blue-gray)

• Type 3: warm, intense, earthy (rust, mustard, olive, camel)

• Type 4: cool, intense, high-contrast (black, white, true red, royal blue, emerald)

Label each type clearly. The goal is for me to evaluate which palette makes my skin look most radiant, my eyes clearest, and shadows the least pronounced.

Use a clean white background, studio lighting, and a side-by-side grid layout.

Why this prompt works:

Most color analysis prompts just say “tell me my colors.” This one forces ChatGPT to generate a visual comparison — four versions of you, each draped in a different seasonal palette. You’re not reading a paragraph trying to imagine what “dusty rose” looks like against your skin. You’re seeing it. The specific DYT color values also prevent the AI from guessing with random shades.

What to look for in your results:

Look at all four images and ask yourself: In which version does my skin look the smoothest? Where do my eyes pop the most? Which one makes me look tired or washed out? That’s your answer.

Prompt 2: Ideal Hairstyle Analysis

Once you know your colors, the next question is always: but what about my hair? This prompt generates a visual grid of different hairstyles on your actual face so you can compare before committing to anything at the salon.

The Prompt:

Using my uploaded portrait, act as a professional hairstylist and create a hairstyle comparison chart.

Generate an image showing me with 8 different hairstyles and lengths arranged in a clean grid layout. Include a mix of:

• Short (pixie, bob, lob)

• Medium (shoulder-length layers, curtain bangs, shag)

• Long (blunt cut, soft layers, face-framing layers)

Keep my natural hair color and face exactly as-is. Do not alter my facial features.

Add short labels beneath each option noting: face shape compatibility, hair texture match, and a suitability rating (great / good / skip).

Use a flat design style, high contrast, and white background. Keep it visual-first — no paragraphs, just clean labels.

Why this prompt works:

The key here is asking for a grid with labels. Without the grid instruction, ChatGPT tends to generate one or two options. Without the labels, you’re left guessing why one style looks better than another. The face shape and texture notes turn this from “here are some haircuts” into actual personalized advice.

Pro tip:

If you love one of the results, follow up with: “Show me option 3 in more detail — front view, side profile, and back view. Include styling notes for my hair texture.” GPT Image 2 handles multi-angle follow-ups really well.

Prompt 3: Personalized Makeup Guide

This is where things get really practical. Instead of watching 45 minutes of YouTube tutorials trying to figure out where your blush goes, this prompt maps everything to your specific face shape.

The Prompt:

Using my uploaded bare-face photo, act as a professional makeup artist and create a personalized makeup analysis graphic.

Generate a visual guide showing 3 complete makeup looks on my face arranged side by side:

1. Everyday Natural — minimal, fresh, 5-minute look

2. Polished Day-to-Night — elevated for dinners, dates, or meetings

3. Editorial / Glam — full impact, bold choices

For each look, use shades and tones that complement my color season (determine this from my skin undertone, eye color, and natural coloring).

Below each look, include short labels for:

• Exact contour and highlight placement for my face shape

• Best blush placement and shade family

• Recommended eyebrow shape

• Lip color family

• Whether I suit warm or cool-toned eyeshadows

Use a clean, modern layout with a white background. Visual-first — icons and short labels only, no long paragraphs.

Why this prompt works:

The three-look structure forces the AI to show you a range instead of one generic face. Including contour and blush placement tied to your face shape makes this genuinely useful — it’s the kind of mapping that makeup artists do in person. And by asking it to determine your color season from the photo, you get shade recommendations that actually match your skin.

What to do with the results:

Screenshot each look. Save them to a folder on your phone called “Makeup Reference.” Next time you’re at Sephora or browsing online, you’ll know exactly which shade families to shop.

Prompt 4: Personal Fashion & Style Direction

The final piece — pulling your color analysis, face shape, and overall vibe into a cohesive style direction. This is where you go from “I know my colors” to “I know exactly what to wear.”

The Prompt:

You are a personal stylist and fashion consultant. Using my uploaded photo, create a comprehensive personal style analysis graphic.

Section 1 — My Style Profile: Determine my color season, body proportions visible in the photo, and overall aesthetic energy. Display this as a clean visual card with icons.

Section 2 — Best Outfit Palettes: Show 4 complete outfit color palettes that work for my coloring, displayed as horizontal color swatches with labels (e.g., “Power Neutral,” “Soft Weekend,” “Evening Out,” “Creative Bold”).

Section 3 — Wardrobe Essentials: Show a visual grid of 6 wardrobe staple items illustrated on or near my image — in the exact colors and styles that suit me. Include items like: a blazer, a great pair of jeans/trousers, a go-to top, a dress, outerwear, and one statement piece.

Section 4 — What to Avoid: Show 3–4 color swatches or style choices that clash with my natural coloring, with a brief “why” label.

Use a modern editorial layout. Magazine-style, not clinical. High-quality, minimal text, icon-driven.

Why this prompt works:

This is the prompt where you really see GPT Image 2’s reasoning ability shine. By asking for multiple sections in one structured layout, you get something that looks like a page ripped from a personal styling lookbook. The “What to Avoid” section is surprisingly useful — most people never think about which colors or styles are actively working against them.

Follow-up prompt to go deeper:

“Now create a capsule wardrobe mood board for me — 15 pieces that all mix and match, using my best color palette. Show the items in a flat-lay grid with labels.”

Tips for Getting Even Better Results

Run them in order. Your color analysis informs everything that comes after. Once ChatGPT knows your palette, you can reference it in follow-up prompts.

Iterate. If a result isn’t quite right, don’t start over. Say something like: “This is close, but adjust the Type 2 palette — the lavender looks too purple. Use a more muted tone.” GPT Image 2 handles incremental edits really well.

Try multiple selfies. Different lighting and angles can shift the results slightly. If you’re unsure about your color season, run the prompt with 2–3 different photos and see which type comes up consistently.

Ask for the reasoning. After any visual prompt, follow up with: “Now explain in text why you chose these colors/styles for me. What did you see in my coloring and features?” The text analysis often surfaces insights the image alone doesn’t.

The Bigger Picture

Here’s what I find fascinating about this trend: a year ago, color analysis was a niche service run by trained consultants. Today, anyone with a selfie and a ChatGPT account can get a version of this for free, in seconds.

Is it perfect? No. AI color analysis can struggle with closely related seasons (like telling the difference between Soft Autumn and Warm Spring), and it can’t replicate the nuance of in-person fabric draping. If you’re serious about building a long-term wardrobe strategy, a professional consultation is still worth it.

But as a starting point? As a way to build awareness of what works for you before you spend money? It’s genuinely powerful — and it didn’t exist six months ago.

Try the prompts. See what surprises you.

Talk soon,

Angel

Part 1 — Desktop

@creativelyangeI need this cat app in my life #newapp #techtok #claudecode #vibecoding #catsoftiktok

Tiktok failed to load.

Enable 3rd party cookies or use another browser

Cat Gatekeeper · Chrome extension

A Japanese developer just released this and it went viral for good reason: when you hit your social media time limit, a chubby orange cat physically plants itself across your screen and refuses to move until your break is over. Free, no ads, no data collected.

How to set it up

1. Open Chrome and go to the Chrome Web Store. Search “Cat Gatekeeper” or use the link below.

2. Click Add to Chrome → Add extension. The cat icon appears in your toolbar.

3. Click the cat icon → set your usage limit (default: 60 min) and your break time (default: 5 min).

4. Open Instagram/TikTok/etc. as normal. The timer only counts while that tab is active — switch tabs and it pauses.

5. When time’s up, the cat appears and blocks all controls. You can’t skip it — you have to wait out the countdown.

Pro tip

Set your first session to 30 minutes and the break to 10. The discomfort of being forced to stop is the whole point — shorter limits = more cat visits = actual habit change.

→ Install Cat Gatekeeper on Chrome Web Store

Part 2 — Mobile

One Sec · iOS & Android

Mobile apps can’t spy on each other by design — only your OS can. One Sec works around this by intercepting apps before they open, forcing a pause screen. You can upload a custom cat photo so your own judgemental feline stares you down.

📱 iOS + Android

How to set it up

1. Download One Sec from the App Store or Google Play. It’s free with a paid tier for more features.

2. Open One Sec → tap Add App → choose Instagram, TikTok, or whichever app you want to gatekeep.

3. In the app settings, tap Custom background and upload a photo of your most unimpressed-looking cat.

4. Set a delay timer — 10 to 30 seconds works well. Every time you tap the app, the cat appears and makes you breathe before you can proceed.

5. Stack it with your phone’s built-in limits: iOS Screen Time or Android Digital Wellbeing. One Sec handles the friction; Screen Time handles the hard cap.

The science bit

The delay doesn’t have to be long. Research on habit loops shows that even a 10-second pause is enough to interrupt the automatic open-without-thinking behavior. The cat is doing real work.

Quick comparison

Cat Gatekeeper

Desktop Chrome only · hard block · actual cat · free · no account

One Sec

Mobile iOS + Android · friction pause · custom cat photo · free tier available

Screen Time

Mobile · hard cap after limit · no cat · built into your phone already

the cat doesn’t care about your excuses
the cat will not be moved
the cat is right

Two platforms that basically every vibe coder and indie builder uses just had major security incidents. If you have projects on Vercel or Lovable, there are specific things you need to do today.

This isn’t fear-mongering. It’s a checklist. Let’s get into it.

🔴 INCIDENT #1: Vercel Was Breached (April 19, 2026)

Vercel — the platform most of us deploy our apps on — confirmed a security breach this week.

Here’s what happened: a third-party AI tool called Context.ai that was used internally by a Vercel employee had its Google OAuth app compromised. The attacker used that access to get into the employee’s Google Workspace account, then pivoted from there into Vercel’s internal systems.

Once inside, they were able to read customer environment variables — but only the ones not marked as “sensitive.” That means API keys, database credentials, signing tokens, anything that was stored without the sensitive flag is potentially in someone else’s hands right now.

Vercel’s CEO Guillermo Rauch confirmed the attack was highly sophisticated and “likely AI-accelerated.” The attacker moved fast and knew exactly where to look.

A hacker group claiming to be ShinyHunters posted on a breach forum saying they have Vercel’s data and are selling it for $2 million. That’s unverified — but the breach itself is confirmed.

✅ What to do right now if you use Vercel:

• Go to your Vercel dashboard → Environment Variables

• Rotate every API key, database password, and token that was NOT marked as “sensitive” — assume they’re already compromised

• After rotating, mark all secrets as “Sensitive” going forward so they’re encrypted at rest

• Redeploy your projects — rotating the variable alone doesn’t update live deployments

• If your team uses Google Workspace, check for a connected OAuth app with ID: 110671459871-30f1spbu0hptbs60cb4vsmv79i7bbvqj.apps.googleusercontent.com — if it’s there, revoke it immediately

🟠 INCIDENT #2: The Lovable Security Flaw (CVE-2025-48757)

This one is older but still unresolved for many builders — and it directly affects your users, not just your credentials.

Lovable’s AI was generating apps with broken database security configs by default. Specifically, it wasn’t properly setting up Row-Level Security (RLS) in Supabase — the database layer that controls who can access what data.

A researcher scanned 1,645 apps built on Lovable. 170 of them were completely open. No login required. Anyone could read your users’ names, emails, payment info, API keys — everything. One app alone had 18,000 user records just sitting there.

When the researcher reported it? Lovable’s official response was: “It’s the user’s responsibility.”

They launched a “security scanner” that only checked whether RLS existed — not whether it was actually configured correctly. False sense of security.

✅ What to do right now if you use Lovable:

• Run Vibe-Scanner (free, open source) against your projects → github.com/SymbioticSec/vibe-scanner

• In your Supabase dashboard, go to Authentication → Policies and verify Row Level Security is enabled AND configured on every table

• Never assume Lovable’s security scan means you’re safe — it doesn’t test whether the policies actually work

• If your app has real users with real data, get a developer to manually audit your Supabase RLS before you go any further

• For future projects: use Lovable for the frontend, handle your backend separately with proper security controls

The Big Picture 🧠

These two incidents aren’t random bad luck. They’re a pattern.

The tools we use to build fast — vibe coding platforms, AI-generated backends, third-party OAuth integrations — all have one thing in common: they abstract away complexity. And sometimes, hidden inside that complexity, is a security assumption that was never validated.

Vercel assumed non-sensitive env vars were actually non-sensitive. Lovable assumed users would catch their own security flaws. Both assumptions were wrong.

Moving fast is good. Shipping without checking is how you expose your users.

15 minutes this week going through those checklists is worth it.

See you next week!

Here’s what happened, what it means, and how to try it yourself.

What is Kimi K2.6?

Kimi K2.6 is a large language model released on April 20, 2026 by Moonshot AI. On paper, the specs are impressive: one trillion total parameters using a Mixture-of-Experts (MoE) architecture, meaning it activates only 32 billion parameters per response rather than all one trillion at once. This is what makes it efficient enough to run without supercomputer-level hardware. It also supports a 256,000-token context window — roughly the length of two or three full novels — and native multimodality (it can process both text and images).

Benchmark results put it in competitive territory with top closed models, particularly for coding and long-horizon agentic tasks — situations where an AI needs to pursue a goal across many steps without stopping to ask for direction. Early testing shows it can sustain thousands of steps in a single session with impressive stability. The weights are published on Hugging Face under a Modified MIT License.

What Are Hermes Agent and OpenClaw?

A model on its own is just a model — you type, it responds. An agent is something different. Agents can take actions: browse the web, write and run code, manage files, call APIs, and chain together complex tasks across multiple steps. Think of it as the difference between asking someone a question and actually delegating a project to them.

Two agent frameworks now work with Kimi K2.6 through Ollama:

• Hermes Agent — built by Nous Research (an American AI lab), Hermes is a self-improving agent framework that ships with 70+ built-in skills, cross-session memory, and automatic skill creation. It hit 100,000 GitHub stars in under two months. It’s MIT-licensed and has been free since it launched.

• OpenClaw — another open-source agent framework available through Ollama, also compatible with Kimi K2.6 from day one.

Both run via Ollama — a tool that makes it easy to download and launch AI models and integrations locally or through its cloud.

How to Try It

If you have Ollama installed, the setup is a single command. To launch Kimi K2.6 with Hermes Agent:

ollama launch hermes --model kimi-k2.6:cloud

Or with OpenClaw:

ollama launch openclaw --model kimi-k2.6:cloud

A quick clarification on “free and local”: the :cloud tag in the command routes the model through Ollama’s cloud infrastructure, not your own machine. This makes it far more accessible — you don’t need powerful hardware — but it does mean the inference is happening on Ollama’s servers. If you want to run it fully locally, the weights are available on Hugging Face, though you’d need serious GPU resources for a model of this size.

Why This Matters

A few things converge here that are worth naming. First, a model genuinely competitive with top closed-source offerings is now freely available and open-weight — meaning anyone can inspect, fine-tune, or deploy it. Second, the integration with agent frameworks like Hermes and OpenClaw through Ollama lowers the barrier to running capable AI agents to basically zero, at least for access. Third, this is another data point in the accelerating trend of Chinese AI labs producing frontier-level open models that put pressure on the entire industry.

Whether you’re someone who wants to experiment with AI agents for the first time, or someone who’s been watching the open-source model space closely, Kimi K2.6 is worth a look. The one-command setup makes it one of the more accessible entry points we’ve seen for this class of model.

Further reading

Kimi K2.6 Tech Blog — Moonshot AI

Hermes Agent on Ollama — Official Docs

AINews: Kimi K2.6 Coverage — Latent Space

Kimi K2.6 on Ollama Library

Thanks for reading. If you found this useful, consider sharing it with someone who’s curious about where AI is heading.

Here’s what you need to know 👇

1. Claude Opus 4.7 — Anthropic’s Best Model is Now Available to Everyone

Anthropic dropped Opus 4.7 this week and it’s a real upgrade. Better at long, complex tasks — think: give it an entire codebase or a 50-page doc and it doesn’t lose the thread. Way better at reading images too, which opens up a whole new set of workflows.

The wild context: this dropped just a week after Anthropic quietly revealed Mythos — a model so powerful they refused to release it publicly because it could exploit vulnerabilities in every single browser and system it tested. Jamie Dimon from JPMorgan literally went on record saying it revealed scary vulnerabilities in global banking. Opus 4.7 is the one you actually get to use. Mythos stays locked in a vault.

👉 Who it’s for: Builders, devs, and anyone doing heavy AI workflows.

2. Adobe Firefly AI Assistant — AI That Works Across All of Creative Cloud

This one is big if you’re in any kind of creative work. Adobe launched an AI assistant that can execute tasks across Photoshop, Premiere, and Illustrator simultaneously — just by describing what you want in plain English.

Not just suggestions. It actually does the work across multiple apps at once. You can even save your multi-step processes as reusable “skills” so you never have to repeat yourself.

👉 Who it’s for: Video editors, designers, content creators — anyone in the Adobe ecosystem.

3. Canva AI 2.0 — Your Design Tool Just Got Agentic

Canva upgraded their AI assistant to actually call tools now. You describe what you want, it builds editable layered designs and automatically pulls in context from Slack, Gmail, and Google Drive. It can also research the web and schedule your content.

This is a big shift — Canva is positioning itself as the AI command center for marketing teams, not just a “make it pretty” tool.

👉 Who it’s for: Marketers, solopreneurs, social media managers.

4. Gemini for Mac — Google’s AI is Now on Your Desktop

Google dropped a standalone Gemini app for macOS. It floats over your screen, reads what’s on it, and answers questions without you having to switch windows. Pulls from your Google Drive and past conversations too.

Think of it like having an AI co-pilot that can see your screen at all times. Great for research, writing, and quick lookups without breaking your flow.

👉 Who it’s for: Anyone deep in the Google ecosystem who wants an always-on desktop AI.

5. GPT-Rosalind — OpenAI Enters Drug Discovery

OpenAI launched a new reasoning model built specifically for life sciences. Biology, drug discovery, translational medicine — it’s trained to reason across all of it.

This one is less “use it today” and more “watch this closely.” AI moving into drug discovery is a huge deal and this is one of the first purpose-built models from a frontier lab for that space.

👉 Who it’s for: Biotech, researchers, investors watching the AI x science space.

The Big Picture 🧠

AI is no longer just about chatbots. This week alone you saw: a model too dangerous to release, creative tools that actually execute workflows, a desktop assistant that sees your screen, and AI entering drug discovery. The pace is not slowing down.

The people winning right now are the ones who stay close to these releases and figure out which tools map to their actual workflow — not just the hype.

That’s what we’re here for.

See you next week!

If someone forwarded this to you, subscribe here
Too many emails? Unsubscribe anytime.

By combining Remotion (a React framework for video) with Claude Code (Anthropic’s agentic coding tool), you can effectively “prompt” a production-ready MP4 into existence. Here is how to set up the ultimate programmatic video stack.

I made this video for my app in 20 seconds using Remotion and Claude Code.

1. The Foundation: Scaffolding Remotion

Before Claude can write your video, you need a playground. Start by creating a fresh Remotion project.

bash

# Create a new Remotion project
npx create-video@latest my-video-gen
cd my-video-gen

# Recommended settings during setup:
# - Template: Blank (best for AI generation)
# - TypeScript: Yes
# - TailwindCSS: Yes (Claude loves styling with Tailwind)

2. The Secret Sauce: Installing “Skills”

Claude Code is smart, but it doesn’t inherently know the nuances of Remotion’s spring(), interpolate(), or Sequence logic. You need to install the official Remotion Agent Skills. This gives Claude a “manual” it can reference while it codes.

bash

# Install the Remotion skills into your project
npx skills add remotion-dev/skills

# Or use Remotion's own built-in command:
npx remotion skills add

This creates a .claude/skills directory in your project. Inside, Claude finds SKILL.md, which contains all the best practices for timing, frame-based animation, and composition structure.

3. Activating the Agent

Now, fire up Claude Code inside your project directory.

bash

claude

Pro-tip: To avoid Claude asking for permission for every file change (which will be many), you can start it with the auto flag for a safer hands-off experience:

bash

claude --auto-approve

Or, if you want to move even faster and trust your setup, you can use --dangerously-skip-permissions — just know it bypasses all safety checks.

4. Prompting Your Video

Instead of writing React components yourself, you can now describe the motion graphics you want.

The Golden Rule: Be specific about frames and easing. Since Remotion is frame-based (30fps or 60fps), Claude performs best when you define animations in frame counts.

Try this prompt inside Claude Code:

“Create a 5-second intro video (150 frames at 30fps). Background: A dark navy gradient. Animation: The text ‘VibeCode’ should scale up from 0 to 1 using a spring animation from frames 0 to 45. Add a subtle floating particle effect in the background using AbsoluteFill.”

Claude will then:

1. Create a new .tsx component

2. Register the composition in Root.tsx

3. Implement the physics-based animations

5. Live Preview & Rendering

While Claude is working, keep the Remotion Studio open in a split-screen terminal. You’ll see the video update in real-time as Claude saves files.

bash

# In a separate terminal tab
npm run dev

Once it looks perfect, tell Claude to render it, or run the command yourself:

bash

# Render to MP4
npx remotion render [composition-id] out/video.mp4

Why this changes everything

This workflow moves you from Editor to Director. You aren’t pixel-pushing; you’re defining intent. Because it’s all code, you can easily swap variables — change a hex code, update a name, or inject an API response — and regenerate the video in seconds.

At first, I made series of invisible instruments, my invisible violin demo resonated with people getting over 40k views across IG and Tiktok. Then I made a series of motion tracking dance apps to improve my dance technique. I ended off the month winning a hackathon for a VR app that turns everyday objects in your room into musical instruments.

In an age of accessible AI tools, the only limit is your imagination and what matters is the quality of your ideas (and your willingness to learn the latest tools)!

What I learned from this whole process is that there’s a big need for people who are not from a traditional coding background, for people who are at the intersection of things. A dancer who codes, a musician who codes, or a skater who codes. These niches allow us to bring in our expertise to create apps that that can enhance our experience in a domain. For example, I’ve played the violin for more than 20 years and that’s how I knew the exact gestures that needed to be programmed into the invisible violin, an the kind of interactions that a user would need to enjoy this product.

All backgrounds and niche interests can bring in some kind of value. I personally find the best apps to be the ones where you are building for yourself. For example recently, I wanted to learn hip pop dance routines in my bedroom, so I build this app where I can slow down and mirror the video for learning a routine. And then I built another one for analyzing my movements and giving me feedback on my techniques.

These apps are most interesting to me because I know there’s a real use case because it has genuinely brought me real value and I use them to improve my dance experience. And I know other dancers also face similar problems.

Diving into your obsessions is a good way to find problems in the world, because you tend to already know a lot about your interests and think more about the problems in that area space.

-

I will be hosting a free XR vibecoding workshop in SF soon, if you want to join the next one, leave a comment below or shoot me a message! #vibecoding

Last week, I had a realization while playing my electric cello: most music software feels like a spreadsheet. So, I spent an hour “vibecoding” a new interface—an invisible instrument that responds to somatic movement rather than mouse clicks.

I didn’t write a single line of CSS. I “curated” it into existence.

The Result

The “How-I-Built-It” Breakdown

To move this fast, I stopped thinking like a developer and started thinking like a Vibe Designer. Here is the stack:

• The Logic: Gemini3 for the spatial audio engine.

• The Flow: I used my Somatic UX framework—testing the “feel” of the interaction before the function.

The Asset Drop

Because you’re a subscriber, I’m giving you the “Skeleton” of this project. Use it to build your own “invisible” tools:

• Subscribe to this substack and I’ll send you the prompt to make this project in my next newsletter!

If you want to learn more

If you want to watch me build stuff like this live and learn the exact “Vibecoding” workflow I used for my 20-apps-in-30-days challenge...

I’m hosting a Pop-Up Masterclass this Saturday. We’re going deep on rapid prototyping, and how to go from scratch into a final prototype that you are happy with.

Sign up here : https://luma.com/7ltz1phz

Sketch on paper. Move to Figma. Hand off static frames to developers. Repeat.

It worked. But it was slow — and honestly, a lot of the time I was producing deliverables that felt more like documentation than design.

This year, everything changed.

The Old Way (And Why It Was Holding Me Back)

The classic design process has a dirty secret: most of the time, you’re not actually designing. You’re translating. Sketches into wireframes. Wireframes into mockups. Mockups into specs. Specs into developer handoff.

Each translation step adds time, introduces miscommunication, and distances you from the thing that actually matters — how it feels to use.

By the time a developer builds your static mockup, you’ve lost two weeks and the design is already stale.

What I Do Now

I skipped the middleman stages entirely.

1. Go straight to high-fidelity. Tools like Stitch let me generate high-fidelity mockups in minutes. Not days. I’m not sketching low-fi ideas anymore — I’m showing clients real, interactive possibilities in our first conversation. The feedback I get is completely different when people can see something real.

2. Vibecode a functional prototype. Instead of a clickable Figma prototype, I build something that actually works. Real interactions, real data, real edge cases. This gets you feedback that means something — not “looks good!” but “wait, what happens when I do this?”

3. Iterate on the real thing. Because the prototype is functional, refinement happens on the live product, not on another layer of abstraction. And when it’s ready, you’re exporting production-ready code — giving developers a head start instead of a spec to interpret.

Where That Saved Time Actually Goes

Here’s the part people don’t talk about: moving faster doesn’t mean doing less. It means you finally have time for the work that actually moves the needle.

That freed-up time goes toward:

• Outcome-driven UX — designing for measurable results, not just beautiful screens

• AI systems design — thinking about trust, transparency, and how people interact with AI-powered products

• Continuous adaptation — because the tools are changing fast, and staying current is the job now

The Practical Shift

If you’re curious about AI but not sure where to start, here’s the simplest reframe: stop thinking of AI as a shortcut and start thinking of it as a thought partner.

Use it to generate options you wouldn’t have considered. Use it to prototype before you’re “ready.” Use it to stress-test your assumptions before a single developer writes a line of code.

Design isn’t about static screens anymore. It’s about going direct to outcome — and AI is the fastest path there.

Are you still starting with a blank canvas? I’d love to know what’s shifted for you.

Subscribe now