Claude Code skills — when to use what

find the on-disk path to THIS chat's raw JSONL transcript

write the CHANGELOG.md + PROJECT_STATE.md entry after changes

end of a session that changed files — tidy, commit, document, report

like finish-session, plus a ~40-min audio course of the session

compact this conversation into a handoff doc for another agent

answer 'what do we know about client/project X' from email/WhatsApp/notes/calls

at the start of work in any repo — loads PROJECT_STATE + recent changelog, shows 'you are here'

when the user invites questions before building — renders the 5-option format

opt-in slash gate to ask clarifying questions before building

pin down domain terminology / ubiquitous language, record an ADR

lock a feature's intent — first stage of the 3-step planning chain

run the whole plan→build chain hands-off — all 3 stages

research the approach, write a checklist + ADRs — stage 2 of planning

throwaway prototype to flesh out a design (terminal app or UI variants)

turn this conversation into a PRD and publish it to the issue tracker

add a new Modal webhook / event endpoint

run the orchestrator locally with a Cloudflare tunnel

deploy execution scripts to Modal cloud

start anything new — scaffolds the Agty project folder + repo-maps standard

create / edit / benchmark a skill

build features/fix bugs test-first (red-green-refactor)

teach the user a new skill or concept, in-workspace

reference for writing skills well — vocabulary + principles

scaffold a project-specific QA skill from a checklist

deep-module design vocabulary — where seams go, testability

scan for deepening opportunities, report, then grill the one you pick

unattended overnight whole-journey QA against the live app

anti-AI-slop finishing pass over a frontend, BEFORE visual-QA

the reusable config-driven Playwright QA engine (usually used indirectly)

multi-layer whole-surface QA/test pass for a project

hand-drawn house-style HTML overview of anything Matt will review

deep multi-agent cloud review of the branch diff (correctness/security/perf)

click-by-click behavioral spec + Expected-Behavior Oracle

block dangerous git commands (push, reset --hard, clean, branch -D)

audit + regenerate a repo's living HTML maps

force the laziest solution that actually works

whole-repo over-engineering audit

code review focused only on over-engineering — what to delete

plan a refactor into tiny safe commits via interview, file as a GH issue

resolve an in-progress git merge/rebase conflict

bring an existing repo up to the repo-maps standard, additively

stand up a human-rating eval lab with LLM-as-judge keep/discard

iterate AI output (prompts/copy) against a frozen eval set + metric

turn code/docs into a queryable knowledge graph; ask 'why did we decide X'

relentless interview to sharpen a plan or design

like grill-me, but writes ADRs + glossary as it goes

stress-test a plan/design by interviewing the user relentlessly

build/query the graph of all Claude Code conversations in a project

on-demand reference: which env var / file / account holds a service's key

Opus-vs-Sonnet + effort-level decision sheet — when to escalate

write prompts at max signal-per-character — use BEFORE composing any prompt

deploy a static build folder to Cloudflare Pages, get a URL

mirror a house-style HTML report into Notion

package content for Dani — HTML on Cloudflare + email + Slack

text → Gemini TTS → Cloudflare → Telegram audio

generate a multi-bot voice-agent config for the coach

build a Hungarian megbízási contract as a Google Doc

manage DocuSeal e-signature workflows from the terminal

embed DocuSeal signing forms / template builder

scrape Google Maps for B2B leads + website enrichment + contacts

create cold-email campaigns in Instantly with A/B testing

issue a Hungarian számla via Számlázz.hu + file it in Drive