Most of what I know about building with AI on this site came from something breaking. Each time, I wrote down what I checked and in what order. Those notes are now 13 Claude Skills, free to download.
A Claude Skill is a folder with a SKILL.md file: instructions Claude loads when a task matches, plus any scripts it needs. Unzip one into ~/.claude/skills/ for Claude Code, or upload the zip in Claude's settings. All of them are MIT licensed.
Where they came from
- My Slack alerts said Applebot had read
/.env.example. It hadn't. A scanner on rented Google Cloud machines was borrowing crawler names, and every request got a 404. That became Crawler or Scanner?. - My assistant's vector index was never used, because a
DISTINCT ONsat above the search. Separately, one oversized page could wipe a whole embedding sync without a single error. That became the pgvector RAG Doctor. - When my API credit ran out, the assistant kept answering from a no-LLM path for days. Making that path fast and honest became Deterministic LLM Fallback. Finding out it had happened at all became the LLM Cost Tracer.
- I graded my own site C+ for AI search. The FAQ only existed as JSON-LD, and 37 of 38 pages shared one meta description. The checklist I fixed it with is the GEO / AI-Search Audit.
- My assistant once called my prototypes "production". The rules that stopped it are in the Honest Evidence Writer.
All 13
Engineering
- pgvector RAG Doctor: why retrieval is slow, empty or wrong, checked in order.
- LLM Cost Tracer: every model call logged with tokens, cost and errors.
- Deterministic LLM Fallback: a circuit breaker and answers that quote instead of generate.
- Crawler or Scanner?: real bot or someone wearing its name.
- Frontend Perf Pass: First Load JS on my homepage went from 178 kB to 129 kB with this.
Search and content
- GEO / AI-Search Audit: can ChatGPT, Claude and Perplexity find and quote you.
- Honest Evidence Writer: case studies with method and limits, and an overclaim check.
- Human Check: how AI-generated a draft sounds, and a rewrite.
- LinkedIn Post Packager: the post, the first comment and a tagged link.
Design and review
- Design Token Enforcer: finds every colour that escapes your theme.
- Ten-Persona Review: your site through ten different visitors' eyes.
- Award-Readiness Audit: a jury-style score with numbered findings.
- CV from Content: a PDF CV rendered from your site's data and theme.
What's not in them
Nothing specific to this site: no keys, no hostnames, none of my content. Where a script was ported from the site's code, like the Human Check scorer or the crawler verifier, it's the same logic as a standalone file with no dependencies.
They aren't clever. Most of the value is the order of the checks and the list of things to stop doing. That's what I wish someone had handed me before each of these broke.