1. Google Docs: agents can now leave comments and suggestions on your Google Docs. This works in Claude Code and Codex. Your org’s tech team can get it working in Claude Cowork and ChatGPT. ⭐

2. Design: Claude Design is really good for websites, presentations, and UIs. ⭐

3. Browser use: Codex CLI can now directly control Chrome and log in to websites via 1Password.

4. Prompt injection: Anthropic have driven attack success rates down to zero. Claude Code now defaults to auto-mode. To repeat from previous newsletters (people keep asking): you can stop worrying about prompt injection when using Claude, ChatGPT or Gemini. Go ahead and connect your agents to Google Drive, Gmail, etc. And enable browser use.

Delegation

5. Life admin: my agents handled 2+ days of life admin last month. End-to-end task completion—not copy-and-paste. They're rivalling a human PA for many things. ⭐

6. One prompt, live website: I one-shotted a website for an academic paper. The result was truly ready to ship with no revisions.

How to focus, despite AI

7. Go nuclear: I start the day with pen and paper. 15+ mins, laptop in bag.

8. Make it easy: I took various measures to make the right actions easy. ⭐

9. Experiment: if I get distracted, an agent yells at me. When I start a chat, the agent looks at my day plan and challenges anything that seems off piste.

Skills

10. Google Doc to Substack: Google Doc → Substack draft with a single prompt, including images, footnotes and tables.

11. Log failure, fix later: log chats that go badly, review them later.

12. YouTube: cut the ads, intro and banter from tutorial videos.

13. Events: I organise dance events with a skill, obviously. Poster, Facebook events, Messenger promotion, comms with the venue.

14. Sentimental: turn your weekly review into a visual keepsake.

Coding

15. Mobile apps: making native apps is now roughly as easy as making PWAs.

16. Sol 5.6: if it’s over-engineering simple changes, switch from “high” to “medium” effort. I also had to disable Compound Engineering.

For power users

17. Usage anxiety: worried about hitting the Fable limit early? Try this one weird trick (a status bar tweak).

18. Find old chats: my ‘cr’ tool finds and resumes agent conversations.

19. AGENTS.md: my latest instructions.

For audiophiles

20. Podcatcher: my personal podcatcher app makes me listen to better stuff. Key thing: good summaries mean better episode choices. Try it!

1. New models: Sol 5.6 and Fable 5 are better at long, complex tasks. I prefer Sol: it's faster, and it writes better. A threshold has passed—try asking for more.

2. Skillz: you can make skills by recording your screen.

Apps I use daily

3. Uplift tips: Pablo's digest of the best AI uplift tweets, delivered via Slack or RSS.

4. Situational awareness: my dashboard summarises the best AI newsletters, and shows a feed of the tweets they link to. It sits on my home screen, where Twitter used to be.

5. Summarisation: my browser extension summarises blog posts, PDFs, YouTube videos and Google Docs. Just press Option + S.

It’s happening

6. What my morning looks like: I describe a typical Thursday, with screenshots. ⭐

7. New Tyler coinage: “AI maniacs are obsessed with working with the latest AI models. They try out new models as soon as they can, spend hours and hours trying to master them, and use them to regulate both their workflows and their personal lives.”

8. Nice framing: knowledge work is becoming more like gardening: create the conditions for growth, don't build plants. Or, in workspeak: you’re a manager, not an IC.

9. AI Twitter: my latest roundup.

It’s useful

10. Image generation: ChatGPT can turn a screenshot into an event poster in two minutes.

11. Screenshot redaction: Codex can redact screenshots, probably reliably.

12. Prompt injection: more evidence that the top labs are winning against prompt injection. Even Simon Willison concedes that their efforts “do appear effective in making these attacks much harder”. As I said last time: you can stop worrying about it when you're using ChatGPT, Claude or Gemini.

For the maniacs

13. Agent on the phone: you can manage Codex sessions from your phone. Works well. The Claude Code version is improving, but still clunky.

14. Parallel sessions: my /pane skill branches work into fresh agent sessions in new terminal panes. ⭐

15. Don't sleep: keep agents running when your MacBook is in your bag. ⭐

16. Codex CLI: computer use is available in terminal.

17. Sol 5.6 concern: there are some reports of egregious “it deleted everything” behaviour, and others describe a lack of caution (e.g. a pull request without prior approval). OpenAI are investigating. If you’re working on production systems, Fable with “auto” mode may be the better choice for July. I’m mostly using Sol, with deterministic guardrails.

18. Visual QA: make agents exhaustively review UIs for visual bugs.

Suggestions? Corrections? Want help? Email me, or tweet.

Want more? Press the button:

Subscribe now

Content warning: I’m mostly learning in public, not writing for an audience. If you find something hard to follow, ask your favourite LLM. ❤️

One thing everyone should know

• Browser use is very good now ⭐

Team adoption

• To raise the floor, push 1-2 key messages per month

• Electricity brought only modest gains for many incumbents

• The 5/2 Cadence: 5 weeks building, 2 weeks investing in AI uplift ⭐

• “Use of AI” should be on weekly 1-1 agendas and performance reviews

Ambitious delegation

• Flight search with Claude Code really works now

• Claude Code does my bookkeeping

• Onboarding a new client for TYPE III AUDIO

Skills

• Skills library for a research team ⭐

• My call summarization prompts

• Turn a messy repo into clean commits ⭐

• Make image

• Make spaced repetition cards (and do chat review)

• Digest of LessWrong comments by Daniel Kokotajlo

Security

• It is totally fine to connect Google Workspace to Claude, ChatGPT and Gemini (and to enable browser use) ⭐

General advice

• Karpathy: remove yourself as the bottleneck

• It is often worth asking: is there a faster approach?

• While you wait, think for yourself

• Tell your AI to flag ambiguous data, not silently guess what it means

Personal apps

• The podcatcher app I’ve always wanted

• Keyboard shortcut to summarise blog posts, PDFs, YouTube videos and Google Docs ⭐

• My break reminder app

• Screen blackout shortcuts to support focus

Terminal workflow

• CLI utility to prompt both Claude Code and Codex ⭐

• Project switcher for Claude Code & Codex ⭐

• Question mode (answer, but don’t do anything!) ⭐

• Remove unwanted new lines when copy-pasting from a narrow Terminal pane

• How to share skills between Claude Code, Codex and Gemini

• Use “Try” for quick experiments

Claude Code & Codex

• How to edit pasted text blocks in Claude Code and Codex

• How to write Figma-style comments on HTML mockups

• How to write Google Docs-style comments on agent plans

• Codex with GPT-5.5 is still better than Claude Code with Opus 4.8, especially at the planning stage

• Enable Claude to test its work in Chrome ⭐

• Does your skill call a bunch of scripts? Consider writing a CLI.

• How to plan agent coding projects

• Keep your Claude Code conversation history ⭐

• Hermes is a big upgrade on OpenClaw

• How to read markdown files on your phone

• Use gdoc to edit Google Docs ⭐

Thoughts? Corrections? Want help? Email me, or tweet.

1. Opus 4.5 & Gemini 3 Pro: Claude and Gemini are now better than ChatGPT for most tasks. Opus 4.5 is the best default model.

2. Coding: this month, it’s Claude Code for everything. Opus 4.5 is a big step up.

⭐ It’s time to try Claude Skills ⭐

Claude Skills with Opus 4.5 are a milestone for mundane utility. To understand, check these examples:

• Create invoices, send them, and verify payment (pulls data from Toggl; makes invoices in Xero; sends via Gmail).

• Review, edit and publish blog post.

• Make an OKR progress report (pulls data from Google, Drive, Calendar, Notion and Gmail).

• Rachel Wollen’s AI Chief of Staff.

Motivated? Learn how skills work, and how to make them.

Want help? Book a call, and we’ll make a skill together.

How to pull data into the Claude web app

3. Notion: Claude can search, create, and edit. Install the connector and configure your system prompt to make it faster. ⭐

4. Call transcripts: Claude can read transcripts directly from Granola, Fireflies and Google Meet. ⭐

5. Read any URL: Claude’s security sandbox limits web access. Use Zapier to work around this, if you understand the risks.

Good to know

6. Google Docs: The sidebar assistant is now powered by Gemini 3 Pro. It is finally not useless.

7. Research: Dwarkesh likes the Google Scholar PDF Reader. Andy Masley thinks that Google Scholar Labs may be a “game changer”.

8. Calendar events: Gemini can create Google Calendar events, and so can Claude Code. ChatGPT and Claude can only read them.

9. Gemini memory: Gemini doesn’t have memory. To get better responses, add “who am I” context to saved info. ⭐

10. Airtable field agents: Send bulk prompts to LLMs and get results in a spreadsheet. Good for bulk classification, extraction, and data enrichment.

11. Evaluation: How to systematically improve your prompts.

12. AI Twitter roundup: If you feel like some wow-scrolling, see this roundup. If you want just one thing, make it Gavin Leech’s review of 2025. ⭐

Coding

13. Notifications: Get notified when Claude Code or Codex finish work. ⭐⭐

14. Front-end design: Claude’s new design skill can one-shot excellent, characterful web design. ⭐

Support AI adoption on your team

15. Playbook: I’ve updated my notes on how to support AI adoption.

16. Internal guidelines: Coefficient Giving shared their internal guidelines on using AI.

17. Robust automations: How to make Zaps that don’t break.

Try this

18. Claude Skills: As above...

19. Reduce friction: Make a browser extension to add keyboard shortcuts to a web app.

20. Fun: Vibe code the Spotify homepage you actually want.

Tell your friends

Please tell your friends about this newsletter. You can tweet this, if you like.

Closing vibe

Gemini 3 Pro is step up for hard questions. But Opus 4.5 is the bigger story for this newsletter: it’s a breakthrough for daily tasks, and for coding.

The “wow” has been strong for me, these past weeks.

Want more? Subscribe.

Please email me. Show me something cool. Ask for advice. Stop me being wrong on the internet.

1. Skills: “Skills” are better than Projects for teaching Claude to do things. ⭐

2. Memory: Claude now gets to know you, much like ChatGPT does. You can review and correct Claude’s memories. And you might want to have ChatGPT tell Claude who you are, or vice versa.

3. Keep using Opus: Opus 4.1 (extended thinking) is sometimes more insightful than Sonnet 4.5. For non-trivial tasks, ask both models.

Good to know

4. Google Docs: How to get Google Doc comments into your LLMs.

5. Transcription: MacWhisper is the best way to transcribe audio and video files. It takes less than 10 seconds for 30 minutes of audio.

6. Wispr Flow: It’s still the best dictation app for most people. But you should enable “Privacy mode” in the settings. Without that, they may use your “anonymised” data “for any business purpose”.

Help your team make better use of AI

7. Guide: It’s time to take AI adoption roughly as seriously as hiring. I’m drafting a playbook. ⭐

8. Workshops: Run weekly LLM sessions using this workshop template.

9. For beginners: 30 days of ChatGPT.

10. Automation: Try the 5x rule—if you do something five times a month, consider automating it.

Coding

11. Coding agents: In October, it was Claude Code for frontend, Codex for everything else.

12. November outlook: Droid CLI, which uses Codex and Claude models, is getting strong reviews. Cursor just shipped a major redesign, and can now test its own work using a web browser.

13. Web apps: Start with the Vercel Supabase Starter for authentication, database, and UI. Pair with AI SDK.

Try this

14. Lapel microphone: Lapel mics let you dictate quietly in a shared office or cafe. They may improve your posture in a private office, and audio quality on calls. ⭐

15. Context: Grant access to your calendar, email and to-do lists to get help with daily plans. Ask: “What’s new in my email? What must I do today? What’s most important this week?” ⭐

16. Make graphs: Take a screenshot, get a graph. They’re really nice. ⭐

17. Build: Got a few minutes? Ask Claude to make a simple software tool or prototype a web app. No coding required.

18. Multi-agent workflows: OpenAI Agent Builder now competes with Lindy, Make and Zapier. It has a much smaller feature set, but it’s easier to play with.

19. ChatGPT: make a FigJam flowchart to design a system, or make a tricky decision.

A very good frame

20. Azeem Azhar: Use LLMs to define tasks, not just do them.

Please reply to this email! Show me something cool. Ask for advice. Stop me being wrong on the internet.

1. Claude: If you’re chatting with Opus, you should usually enable extended thinking. Many people haven't noticed this option, which warrants a 🤦 emoji.

2. Claude privacy: For personal accounts, “train on my data” is default on when you accept their new terms & conditions. Turn it off in Settings > Privacy.

3. Google Docs: You can now listen to Google Docs.

How memory works

1. ChatGPT: ChatGPT gets to know you, a bit like a human would. My note explains how it works, and how to control it.

2. Claude: Claude does not have memory, but it can reference past chats. You'll always see when it does this, and normally you must explicitly request it.

Apps I like

1. Claude Code: It has improved a lot since May. Here are some impressive examples.

2. Codex CLI: With the release of GPT-5, some Twitter vibes suggest that it’s a better coding agent than Claude Code. My early impressions are consistent with this—faster, smarter, cheaper. For now, the Claude Code UI remains better, and that matters.

Apps I still like

Mentioned before, but worth highlighting again, now that I've learnt more.

1. Airtable: Got a list of names and email addresses? Airtable AI fields can get job titles and LinkedIn URLs from the web. Their sidebar assistant is good, especially if you're new to Airtable.

2. Granola: I've written a detailed review of Granola for call transcription.

Inspiration

1. Good framing: “AI is great when I’m a noob, but not when I’m a world expert”.

2. Roundup: Collection of “How I use AI” posts and podcasts.

Try this

1. Shortcuts: Make keyboard shortcuts for common prompts. ⭐

2. Prompting: Ask your LLM to help you write prompts. ⭐

3. Computer use: ChatGPT Agent can book you a hotel, find the best flights and record invoice payments in Xero. I will actually use this, which feels like a milestone.

Suggestions? Corrections? Want help? Email me, or tweet.

1. Terminal: Warp is a terminal you can talk to. Developers should try it.

2. Narration: ElevenReader may be the least bad way to listen to articles and Google Docs. Here's a quick way to make complex texts more listenable.

3. Shortcuts: BetterChat now works with Claude.

Want me to find you the best app for X? Know something I might like? Email me.

Good to know

1. Airtable: "field agents" can analyse documents and get information from the web. So if you have a spreadsheet of CVs or website URLs, you can use LLMs to assess them right within the sheet.

2. Claude & Asana: You can create Asana tasks from text or screenshots. With a bit of config, this works well on desktop and mobile.

3. Todoist: The desktop app can create tasks with voice.

4. Raycast: Raycast AI enables fast text manipulation, which is great for proofreading.

Try this

1. Planning: Plan your week with Claude. ⭐

2. Workflow: Collect context in Google Docs. ⭐

3. Google Docs: Enable markdown to preserve links and formatting when copy-pasting into LLMs.

Get a free AI uplift consultation

For the next two weeks, I'm offering free AI uplift consultations.

I'll ask you about your role, and your current use of LLMs and other AI tools. Then, I'll offer pointers and suggestions to help you get more useful work out of AI.

Book a slot here.

Feedback? Suggestions? Want help? Email me, or tweet.

1. Voice input: Wispr Flow. To see how, watch this two-minute screencast.

2. Shortcuts: I've released BetterChat, which adds keyboard shortcuts to ChatGPT.

3. Call transcription: Granola for video calls, and the Google Pixel "Recorder" app for in-person meetings. Granola summaries are fine, but o3 does it better. I'm currently testing Circleback.

4. ChatGPT o3: Currently the best model for (most of) my purposes. I routinely send identical prompts to o3, Claude 4 Opus, Gemini 2.5 Pro and Grok (now Grok 4 Heavy). At least two-thirds of the time, o3 wins. But it's still worth sending important stuff to several models.

Good to know

1. Web search: ChatGPT o3 has much better web search than Claude 4 Opus and Gemini 2.5 Pro.

2. Google Sheets: There's an =AI() function. E.g. =AI("categorise sentiment as positive or negative", A2). Docs here.

3. Data analysis: clean CSVs first and use o3 or Gemini. Do not use Claude.

Try this

1. Context: Make an "about me" doc, to tell the AI who you are. This enables much better chats about planning, personal challenges, life dreams, and much else. ⭐

2. Nutrition: Low-effort diet tracking: take a photo, get the numbers (and some great advice).

3. Vibe-coding: Make websites better with a personal browser extension.

4. Voice input: Mentioned above. But seriously: if you haven't tried it recently, do it now. ⭐

Fun

1. Suno AI: Glue yourself to the Autobahn.

Want more? Subscribe.

Questions? Corrections? Want help? Email me, or tweet.