Claude Code
How I go from idea to working product every time
Build Your First Website
The idiot's guide. I taught a mate who'd never properly used a Mac to go from zero installed to a real website, live on the internet with a link he could send his dad — in an afternoon, entirely by talking to Claude. A visual, click-through walkthrough: every step from installing the editor to deploying on Vercel, with a "what goes wrong" box for each bit that trips everyone up.
The LLM Council
Turn one question into five expert opinions, then a single clear verdict. Five advisors reason from different angles, peer-review each other, and a chairman tells you what to actually do. Claude is the world's most convincing yes man — this is how you make it argue back. One prompt installs it from my GitHub.
Replace Line Items, Not Job Descriptions
When people start with AI at work they picture it doing the whole job — the excited hand it everything, the worried fear it takes everything, same mistake both ways. Your job is a bundle; AI is terrible at being you and brilliant at the tasks you repeat. One paste-in prompt interviews you about your week, finds your first line item, and helps you hand it over.
The Canary Trick
Get your AI to call you by your name in every reply — not as a gimmick, but as an early warning system. One line in your settings, and the moment the name disappears you know the chat is degrading and the rest of your instructions are slipping too. Works in Claude, ChatGPT and Gemini, with the CLAUDE.md version for builders. Borrowed from Van Halen's brown M&Ms, inspired by Peter Steinberger.
Which Model For Which Job
New models keep appearing in your dropdown and nobody explains which to use. The honest answer is that the generation barely matters anymore — the size does. Think of them as people at a firm, point the senior partner at almost everything, and know the two cases where that rule bends. Works the same in Claude, ChatGPT and Copilot.
Claude For Chrome: Set It Up
Anthropic's official browser extension puts Claude inside Chrome, logged in as you, so it can act on the stuff behind a login that a normal integration can't reach: filling portal forms, staying on top of the messages you've left on read, running tasks on a schedule. How to install it, a safe first task to try, and the honest bit on what not to trust it with.
The AI Autonomy Ladder
Where's your line with AI? Not one line for everything, drawn on a feeling, but three levels a task can sit at: in the loop (you drive), over the loop (it drafts, you approve), out of the loop (it just runs). The trick most people miss is splitting a job across all three and letting the AI sort the work itself. Includes a prompt that maps your own tasks onto the ladder, built to take into work.
Can You Trust What AI Gave You?
The billion dollar question, answered with one two-second test: did you witness what the AI is working from, or not? Witnessed (a call you were on), you're the check and a read is enough. Not witnessed (a report you never read), a wrong answer looks exactly like a right one, so you have to build the checking in. The trap is that we trust it hardest exactly where we can check it least. Includes the three checks for when you can't be the test, and a prompt that sorts your own AI tasks.
What Millions of Claude Conversations Reveal
Anthropic studied how people actually use AI, and it turns out we're all creatures of habit: sleep advice at 5am, the news at 7, emails nine to four, recipes at six, and the higher-paid still working at midnight. My write-up of the daily rhythms with the report's own charts, how they know all this without any human reading your chats, and the link to the full study.
Get Your Business AI Ready
95% of corporate AI projects deliver no measurable return, and the reason I kept seeing from the inside is the least glamorous one: there's nothing for the AI to read. The five switches to flip before you spend a pound on tools — transcription on, one findable home for documents, access sorted, policies actually written down, simple formats — plus a prompt that interviews you and builds the capture plan for your business.
Make AI Admit It Doesn't Know
One line that stops more hallucinations than anything else I do: give the model permission to say "I don't know". The exact line to paste, how to make it permanent in Claude, ChatGPT and Claude Code, and the reason it works (the same reason your kid blames the cat). Honest bit included: it won't catch things the model wrongly believes — it removes the pressure to invent, nothing more.
The Privacy Toggle: You're Half Right
Turning off training on Claude and ChatGPT genuinely works — it's a legal commitment with regulator teeth, and it's the difference between your chats living for 30 days or 5 years. But it only answers half the question: your data still lives on US servers whatever the toggle says. Where the switches are, what never goes in a personal chat, and the Europe options if your agreements care.
Have Fable Interview You
Don't fire questions at the smartest model before it leaves your plan — flip it. One paste-as-is prompt: Fable interviews you about your life and work, one question at a time, then writes a two-half document you keep — your situation distilled, plus its honest read on what matters and what it would do first. Works on any model if you've missed the window; the second prompt hands the document to every future chat.
Context Engineering, Not Prompt Engineering
The models change every few months; how you communicate with them is the skill that carries over. Why the wording was never the bottleneck, the two questions to ask before every prompt, an interview prompt that makes the model pull the context out of your head, and the case for dictating instead of typing — with the free tool I built for it.
The Fable Window: Three Dos, Three Don'ts
You've got the smartest model until Sunday 12 July, and most of that window gets spent on work that doesn't need it. The three things worth the window and the three that waste it — each with the exact prompt or claude.md line to paste, plus the reasoning: why small edits don't need Fable, why the security review moved to Opus, and why files beat chat. Links on to the deeper guide for every one.
The claude.md Files for Non-Technical Professionals
What made Fable special was never just the size of the model — it's the way it thinks, and a way of thinking can be written down. Three claude.md files written for non-technical professionals: thinking rules, not coding rules. One paste and Claude sets everything up — or interviews you and builds your files around your actual work. For a lawyer, an accountant, or whoever you are.
The Five Skills Fable Wrote Before It Left
A skill is a manual: a file where the model writes down exactly how it does a job, and every model after just follows it. Before Fable left the standard plans I had it write the five worth keeping — the security sweep, the setup, the build planner, the honest advisor, the bug hunter. Install all five in one paste, plus the prompt to commission your own. Each one ends with its honest limits.
The Three Tiers of CLAUDE.md
Claude reads your CLAUDE.md files before every single prompt — its standing orders. Most people know about one; there are actually three tiers: folder, project, and your whole computer. What belongs at each level, a paste-ready starter template for all three, and the distill prompt that gets the smartest model you have to write yours — so its way of thinking stays even when the model goes.
10 Things to Do With Fable Before It Leaves Your Plan
Fable 5 stops being included in Pro, Max and Team plans on Sunday 12 July — after that it's usage credits on top of your subscription. The ten best things to run through the smartest model Anthropic have ever shipped before the meter starts, each with the exact prompt to paste. Increasingly niche on purpose, and the last four keep paying you back after the window shuts.
Make Your App Stop Looking AI-Generated
There's a dead giveaway that an app was vibe-coded: the same fonts, the same grey cards, the same layout as everyone else's. That's Claude designing from memory. The fix is handing it parts to copy — 21st.dev is a library of thousands of properly designed React components, and this is exactly how to hand one to Claude. Plus the honest bits: quality varies, check the licence, and don't trade generic for jumble.
The Effort Dial
The magic thinking words — think hard, ultrathink — are folklore now. The real control is /effort: five levels from low to max that tell Claude how hard to try. Which level fits which job (and which one quietly drains your limit), what ultracode actually is, and an honest note on when the dial makes no difference at all.
10 Claude Code Commands You're Not Using
Anthropic have quietly built 75+ slash commands into Claude Code — you just type a slash. Most people use about three. The ten worth knowing: /init to give Claude a memory of your project, /context to see what's eating your tokens, /rewind to undo a change, /security-review before you push, /loop, /deep-research, /code-review --fix and more. Use all ten and you're in the top 1%.
The 4 Modes in Claude Code
Most people use Claude Code in one mode the whole time — the slow one, where it stops to ask permission for everything. There are four, and you flip between them with Shift + Tab: ask-before-edits, accept-edits, plan, and auto. What each one does, when to reach for it, the bits people get wrong (no, accept-edits doesn't say yes to everything), and the dangerous fifth setting you can't cycle to.
Turn Anything Into a Podcast
NotebookLM turns whatever you want to learn into a two-person podcast you can listen to on a walk — and most people stop at the default. The bit worth knowing: in Customise you tell the two hosts who to be. Have them rehearse a pitch (one plays the CEO you're pitching to), teach you something (one lecturer, you the student), or debate a decision both ways. The setup with screenshots, plus copy-paste persona prompts.
Make Your AI Stop Agreeing With You
AI is agreeable by default — it's trained on the answers people rated highest (the flattering ones), and it reads your wording for the answer you're hoping for. Two paste-ready instructions — one for Claude, one for ChatGPT — that sit above every chat and make it push back, point out the flaws, and tell you when you're wrong. Where each one goes, why it works, and an honest note on what it won't fix.
Catch the Holes Before You Ship
Vibe-coded apps get hacked on the same handful of mistakes — exposed keys, missing auth, injection. Claude Code's /security-review scans the changes on your branch for exactly those, reports each with a severity and a fix, then fixes them on your say-so. The command, a deeper copy-paste prompt, how it works, and an honest note on what it won't catch.
How to Set Up Connectors
Most people use AI for advice, then go and do the work themselves — clicking between calendar, inbox and docs, gathering context by hand. Connectors plug your real tools into Claude (or ChatGPT) so it reaches in and does the gathering itself. The five-step setup, what to ask first, and an honest note on access and permissions.
4 Loops That Have Nothing to Do With Code
Loops aren't just for engineers. Four that quietly run my week — an evening wrap-up that updates my CRM from meeting transcripts, a LinkedIn chase, a brief before every meeting, and my watch pulling in sleep and recovery. Paste-ready setups for the first three; the watch one's bespoke.
The Two Types of Loop
Before you start a loop, know which kind you're running. Deterministic loops hit a number they can measure (under 50ms, tests pass). LLM-as-judge loops grade fuzzy goals against a rubric ("until the code's clean"). When to use each, why your loop never finishes, and what a rubric actually is.
7 Ways to Actually Use Loops
Everyone talks about loops. Nobody says what to point one at. Seven named, practical loops — the speed loop, the overnight docs sweep, refactor till happy, logging, the nightly error sweep, an SEO + GEO audit, the eval loop — each a paste-ready prompt that gets Claude to set the whole loop up for you. Inspired by Matthew Berman.
Put Claude's Juniors to Work
By default one Claude does every part of a job personally, on the biggest model you've got — the senior partner renaming files. Subagents fix that: junior copies on smaller models take the grunt work and the heavy reading, reviewed before it lands. Two paste-ready lines make it standing behaviour — one for your bill, one for your chat.
The Context Window
Every message you send, Claude re-reads the entire conversation to reply — the system prompt, the chat history, all of it. So long chats get slower, pricier, and more confused as they go. What the window is, what a token is, and the handoff prompt that moves your work to a fresh chat without losing anything.
Stop Telling Claude It's an Expert
Almost every prompting guide tells you to open with "you are a world-class expert." Controlled studies say it doesn't improve accuracy — and can lower it, while making Claude sound more confident than it should. That, plus nine other ways to actually prompt it better.
The Verify Loop
Everyone's saying stop prompting and write loops. The loop is the easy part — a while-loop with nothing checking its work just agrees with itself twenty times. This skill is the loop done right: it runs one task, actually verifies it, commits on green, and stops itself. Built on Boris Cherny's loop engineering.
100+ AI Models, Free
DeepSeek, Kimi, Qwen, GLM and more — the best open models in the world, hosted by NVIDIA and free through one API key. It's OpenAI-compatible, so it drops into whatever you already build with. Grab a key, paste one prompt into Claude, done.
The Compliance Audit
Your vibe-coded product is probably illegal — not on purpose. This skill reads your repo and checks it against GDPR and the EU AI Act, then hands back a ranked list of exactly what to fix. Privacy policy, data deletion, AI disclosure, third parties. Inspired by Flagged.
The 3-Step AI Build System
Every AI project I build starts the same way. Not with a prompt. Not with a tool. With a notepad and five minutes of thinking. Pen & Paper, Plan Mode, Auto-Accept.
The Coding Guidelines CLAUDE.md
A drop-in CLAUDE.md with four rules that stop AI from overcomplicating, over-editing, and guessing. Think first, keep it simple, change only what's asked, prove it works. Copy it or let Claude install it.
GPT Prompt Engineering
Two drop-in Claude skills that make Claude write GPT prompts the way OpenAI actually recommends — one for GPT-4.1, one for GPT-5. Copy or download and drop into your project.
User Story Creation
Turn a feature idea into a detailed GitHub issue. Claude reads the codebase, plans the implementation, and creates a structured story — ready to build.
Teach AI to Do a Task Your Way
AI rarely does the same job the same way twice, so you end up re-explaining what you want every time. Teach it once instead: get it right, save the recipe, and it does the job your way from then on — no code. The no-code way for ChatGPT or Claude, plus the more powerful skill version if you use Claude Code.
How to Host on Vercel
Get a project from your laptop to a live URL — told as the prompts you give Claude, plus the two moments you take the wheel yourself.
How to Make Videos with Veo
Google's Veo makes short cinematic clips — but the prompt is everything. Hand Claude a template, chat through your idea, and paste its prompt into Flow. Plus which model to pick.
Paxel
A new YC tool that reads your AI coding transcripts and builds a profile of how you actually work — your archetype, your habits, your patterns. Copy one prompt, paste it into Claude, and it sets you up.