These are the working instructions I give Claude Code, Anthropic's AI coding tool, to run jobs I used to do by hand: checking a proposal before it goes out, testing whether an AI answer holds up, planning the week from what actually happened. On GitHub, no sign-up, nothing behind a form.
The repo
Each one exists because I was doing the job badly by hand and wrote it down. Each links to its folder on GitHub: a write-up of the method and what it cost me to learn, plus, for most of them, a set of instruction files (a "skill") you drop into Claude Code so it runs the job the same way every time. Where I know the limits, I say so.
Never used GitHub? Each link opens a page of files with a plain-English explanation at the top. You can read all of them without installing anything. Installing takes about five minutes and needs Claude Code.
Reads anything about to leave your hands through the eyes of the person receiving it, and reports how it lands at three depths: a five-second skim, a full read, and the between-the-lines interpretation. It flags tone friction and missing pieces, fixes them, and loops until the draft is at "I'd send this as-is".
Useful forClient emails, proposals, decks, posts, anything public-facing where you only get one shot at the first read.
Copy-paste prompts and a skill for testing whether an AI result can be relied on: a report, an analysis, a research file, a summary, a calculation. AI output reads well on the first pass whether or not it is right, which is what makes it dangerous. This is the pass that checks it.
Useful forAny AI output you are about to act on, send to a client, or put a number in front of.
Reads your last week from sent email, calendar and notes, checks it against a short file of your goals, and hands back three to five priorities for the week ahead. Each gets a reason, a next action, a definition of done and a time block, with exactly one marked as the week's number one.
Useful forWeekly planning that starts from the evidence of your last week. I have run a version of this every Sunday since July 2026.
Two skills for the kind of page you send as a link after a cold email, where the reader should understand the offer and act without a call. The first builds the page with you stage by stage. The second reviews any page the way a first-time visitor sees it.
Useful forOffer pages, waitlists and demand tests. Written after one of my own pages went through twenty versions in two days, most of them fixing things a better start would have settled.
A method, research notes, templates and a skill for the file that tells Claude Code how to work with you, the CLAUDE.md file. These files accumulate: a rule here, a runbook there, project context, footguns turned into "never do this again" lines, until the assistant starts ignoring the whole thing.
Useful forAnyone a few months in whose setup has quietly stopped behaving the way the instructions say it should.
Takes a company name and produces a sourced research archive plus one self-contained interactive page that opens on what to do, backed by a dated timeline, go-to-market, competition and a deep dive. Separate AI agents research and check the work, and the page states how far to trust it.
Useful forCompetitive profiles, prospect research, and any deep dive where the error rate matters as much as the findings.
Two skills that run as a pipeline. The first reads your posts and builds a profile of how you write, where every trait is tied to a quote from you and how often it occurs. The second turns that profile into a guideline an AI writes from, with a built-in check against the usual AI tells.
Useful forGetting drafts that sound like you. Every claim about your voice comes with the quote it came from, so you can check it instead of taking it on faith.
Hand your brand's design files to an AI agent and the output comes back looking almost right, with no error anywhere to explain why. This is the write-up of what actually breaks: fonts that never ship, preview cards that render nothing, and a compiled stylesheet missing the colours your own brand doc promises. Everything reports success and the output is still wrong.
Useful forAnyone handing a design system to an AI agent, and anyone who has to explain why the result looks subtly off.
Write-up only. This one has no installable skill.
A skill that sets up your home on HomeExchange or Kindred. It reads a folder of your home's info and photos, asks only for what is missing, writes the title and description, uploads and orders the photos, and fills in the platform's setup in your own browser where you are already logged in.
Useful forThe clearest example of the whole pattern. A tedious job I did once by hand, written down, so now it runs in one session. The same shape works for any repetitive setup task in a business.
All nine playbooks, the skills that go with them, and whatever gets added next. Public, MIT licensed, no sign-up.
Along the way
I write up what I am building and what broke while building it, usually before it is tidy enough to become a playbook. Field notes, while the thing is still messy.
The playbooks are how I work. Book 30 minutes and we will find the one move worth making in yours.
Book a 30-minute sessionNo pitch. If I am not the right person for it, I will say so and point you somewhere better.