create-verification-skill · git:20261108.ac1df7f · 2026-11-08 · sha256 184d4240c4715926
create-verification-skill git:20261108.ac1df7fA
Immutable. This exact content never changes and is served at /api/v1/blob/184d4240c4715926.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_uj6d3ruqrqdv62fa/184d4240c4715926b49858813afdc028b70f9bd278296410151b9d99b23395b9.json
Then /plugin install create-verification-skill@markdownregistry-uj6d3ruqrqdv62fa in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
---
name: create-verification-skill
description: "Generate the project's own verify skill: build/test/lint commands discovered by running them, a feature map of what each covers, manual checks, and an earned first verdict. Run once per repo."
aliases: []
metadata:
credits:
verification-skill-generator: pstack
author: Lauren Tan
url: "https://github.com/cursor/plugins/tree/main/pstack"
port: pstack-claude
port-author: Michael Denyer
port-url: "https://github.com/michael-denyer/pstack-claude"
---
# Create Verification Skill
The bundled `verify` skill demands evidence but cannot know this repo's
commands. This skill generates the project-local `verify` skill that makes
verification concrete — by discovering the commands through execution, not
by guessing from a README.
## When to use
- a repo has no `.claude/skills/verify/` and work will be verified here
more than once
- the generic verification flow keeps guessing commands wrong for this repo
## Workflow
1. **Discover by running.** For each candidate surface — build, test, lint,
typecheck, format, dev server, e2e — find the command (package scripts,
Makefile, CI config) and **run it**. Record the real invocation, the
expected passing output shape, the runtime, and the known gotchas
(env vars, seed data, order dependence). A command you have not run is a
hypothesis, not a check. Write-mode formatters (`prettier --write`,
`eslint --fix`) mutate tracked files — record the read-only check form
(`prettier --check`, `eslint --check`) as the verification command, never
the mutating one.
2. **Record the manual surface.** What cannot be checked by command here:
the pages or flows a human (or `run`/`visual-verdict`) exercises, and
what "looks right" means for each.
3. **Write the feature map.** A table of what each check actually covers —
which entry points, which layers, what it cannot see. The map is the
skill's honesty: verification strength is knowable, not assumed.
4. **Generate** `.claude/skills/verify/SKILL.md` containing: the exact
commands with expected outcomes, the manual checks, the feature map, and
the escalation path (which failure means what, who owns it). Keep it
short enough that a fresh session reads it whole. If a verify skill
already exists at the canonical root (`.omc/skills/verify/SKILL.md` or
`.claude/skills/verify/SKILL.md`), respect it: update its commands and
feature map in place rather than overwriting, and say what changed.
5. **Prove it.** Run the generated skill's own flow once end to end on a
real change and record the result through the ledger's locked writer —
`appendLedgerEntry` and `defaultLedgerPath` from the installed package's
`dist/features/verification-ledger/index.js` (`babysit`'s ledger step
demonstrates the exact invocation); never append raw rows to
`.omc/state/verification-ledger.tsv` — the writer owns header creation,
locking, and cell sanitization. The verdict is the one the flow actually
earned: a real UI exercised is `live-ui-verified`, a test suite run is
`unit-test-verified`, only build/lint/typecheck passed is
`type-check-only`, and nothing runnable is no positive row at all.
## Rules
- Every command in the generated skill was executed this session, or is
marked unverified with the reason.
- The generated skill verifies; it never fixes. Fixes route to `execute`.
- When the repo's commands change shape, `maintain-verification-skill`
re-syncs this artifact.