ai

Shared Claude Code skills, collected from across my repos so they live in one place instead of being copied between them.

Using them

Link this repo into a session alongside whatever you are actually working on. Claude Code loads a linked repo's .claude/skills/, so everything below becomes available as /<skill-name> without any install step or per-repo config.

That is the whole setup — the layout here exists to be discovered:

.claude/skills/<name>/SKILL.md    # one directory per skill
.claude/agents/<name>.md          # subagents the workflow skills fan out to

Watch for name collisions

Several of these skills may also exist in the repo they came from. Link this repo next to one that carries its own check, plan, typescript, or tests skill and you have two of each. Only one copy of a given name wins, and which one is not worth guessing at.

So this is most useful linked next to a repo that does not already carry these. To get the clean version everywhere, the duplicates eventually have to come out of the source repos; nothing here requires that first.

There is deliberately no CLAUDE.md

A linked repo's CLAUDE.md loads too, and this one has no business telling another repo how to build. Keeping it absent means linking this repo adds skills and nothing else.

Skills

Workflow

Skill	What it does
check	Fans the branch's diff against main out to read-only checker subagents that try to break it, then reduces the findings to a plan.
check-iterate	Runs that as a loop — checker collects, verifier refutes, the primary agent fixes and pins each fix with a regression test — until a pass comes back clean or seven loops.
plan	Writes plan files in the .claude/plans/<slug>.plan.md stages schema.
plan-execute	Executes a plan file, runs the plan's own verification commands, drives the check loop over the result, then publishes the run through session-report.
feature	Takes requirements to merged PRs on Claude Code cloud — plans the work, gives each stage its own cloud session, branch, and PR, reviews each PR with fresh-context adversarial rounds, and merges when the gate passes. Stops once, for plan acceptance.
session-report	Publishes what the session did as a linked Claude artifact: what was asked, what changed, what was verified, every finding with its verdict and fix, and what needs a human. Leaves the link on the PR when there is one.
pr-review	Reviews one PR (or the current branch, with --branch) against a prompt you supply. Posts line-anchored comments, capped at a COMMENT review.
pr-triage	Reads the whole open PR queue and publishes one page rating every PR Low → Critical, with a "start here" list.
async-daily	Writes an async standup grouped by pull request — what merged, what is still open, plus local work that is not a PR yet. Covers every repo you touched, not just the ones checked out.
debug	Reproduce, localise, hypothesise, fix, prove: root-causes a bug or flaky test and leaves a regression test behind. --no-fix stops at the diagnosis.
qa	Runs a QA pass on a feature in a real browser — Claude in Chrome where the session has it, a Playwright fallback where it does not — reproduces every failure before reporting it, and publishes a report with a verdict, findings, evidence, and what was not covered.

check, check-iterate, and plan-execute stage their work; none of them commit. debug edits but never stages or commits. async-daily, session-report, and qa are read-only in the repo; session-report publishes an artifact and may comment its link on the branch's PR, and qa publishes one too. qa is the only skill here that touches something outside the repo — a running app — so it confirms the target with you before it starts and refuses destructive checks against production. feature is the exception to all of it — it commits, pushes, opens PRs, and merges them, which is why it is the only skill here that stops for an explicit human approval before it does anything.

Agents

Agent	Role
checker	Read-only collector. Assumes the branch's diff is broken and hunts for evidence: a concrete input, a caller that did not adapt, a test that would still pass if the code were wrong. Grounds itself in the repo's own agent guidance first so a settled convention is never a finding.
pr-adversary	Read-only reviewer for one feature review round. Same mandate as checker, but it reads a pull request's diff through the GitHub API at a fixed head SHA rather than a local checkout, and it also polices the things only a fan-out run can get wrong: files outside the slice's lane, drift from the plan, and edits that weaken the very gates the run merges on.
verifier	Read-only refuter. Takes another reviewer's findings and defaults to REFUTED, confirming only what it can trace itself. The two agents have opposite mandates on purpose: a finding has to survive both before anyone edits.
run-reconciler	Read-only state collector for one feature wake. Checks every ledger row against GitHub and the worker sessions and returns only the rows that drifted, so raw PR and CI payloads never fill the long-lived lead's context.

All four run on Sonnet 5.5, the reviewers at medium effort and run-reconciler at low, since collection is wide and parallel and the judgment calls stay with the orchestrator.

Conventions

Skill	What it covers
typescript	Typing, Result-based error handling, ESM conventions, naming, JSDoc.
tests	Vitest and Playwright practice, and running suites scoped through the repo's task runner.
comments	Why-not-what comments, JSDoc, and the global-CSS exception.
vue-components	Composition API structure, typed props and emits, Tailwind, accessibility.
openapi-glossary	Consistent OpenAPI terminology.

Where these came from

Most of the workflow and convention skills were collected from other repos; async-daily, debug, qa, session-report, feature, and the verifier and pr-adversary agents were written here. typescript and tests existed in several divergent copies — the richest of each was taken and the repo-specific commands generalised.

Repo-specific paths, package names, and task-runner commands are either removed or stated with the condition under which they apply. pr-review resolves the repository from origin rather than assuming one, and reads whatever agent guidance the repo has instead of a fixed list of files.

What is deliberately not here

Skills that only work inside one repo stay there — anything tied to a specific layout, fixture set, or toolchain. session-report replaced one such repo-specific report skill with a rewrite that depends only on git and the Artifact tool, and is renamed so it never collides with the original. The bar for adding something here is that it would work unchanged in a repo that has never seen it.

Adding a skill

1. .claude/skills/<name>/SKILL.md, with any references and scripts beside it.
2. Write the description for triggering — what it does and when to reach for it, in the words someone would actually use.
3. Strip repo-specific paths, package names, and commands, or state the condition under which they apply.
4. python3 scripts/validate.py.

The validator checks each skill's frontmatter name matches its directory, that it has a description within Claude Code's listing limit, that every <skill dir>/… pointer and relative Markdown link resolves, that no Cursor mdc: links survived the move, that allowed-tools names the subagent tool as Agent (there is no Task alias, so a Task entry pre-approves nothing), that every subagent_type a skill spawns has an agent file, and that no bundled file sits there unreferenced. Agents get the same treatment against the fields the sub-agent docs define (model, effort, color).

Adding an agent

1. .claude/agents/<name>.md with name, description, tools, and a model.
2. Give it one mandate and state it in the first paragraph. checker assumes the code is broken; verifier assumes the finding is. An agent that hedges between two stances does neither job well.
3. Define its output format in the file, so the skill that spawns it can parse the result without restating the schema in every prompt.
4. python3 scripts/validate.py.