heznpc Turn completed implementation into reviewable evidence
When the requested feature is implemented, its configured completion proof is
part of the handoff; the user need not separately request a screenshot. First
inspect take-a-repo.config.js and the actual commands before executing them.
take-a-repo inspect <repo> --json only discovers capabilities; it never runs
repository code. A missing config requires product-specific intent, not guessed
claims. Read the package's evidence contract to author
one when release-proof creation is within the user's request.
Use the product's existing execution tool: Playwright for browser interactions, a repo-owned Appium/Maestro/fastlane/native runner for native UI, executable checks for a CLI, actual request/response assertions for an API. Do not build a web mock to stand in for a native or nonvisual product. The shared engine accepts their evidence; it does not implement every platform's automation driver.
Run take-a-repo <repo> --json (or the source checkout's absolute
bin/take-a-repo.js with Node). Match every product claim to an observed check,
including meaningful failure behavior. A screenshot proves appearance, not a
successful transaction. Imported media and checks remain unverified until the
product is actually exercised. Do not replace a failed check with pass to make
the delivery green.
Read the returned run report and its actions. Fix the responsible producer,
assertion, story or rendering recipe, then recapture. Increment --attempt; stop
and explain when the configured budget is exhausted, a required permission or
external dependency is missing, or fixing the issue would expand the request.
Scoped evidence runs are diagnostic only; run the full config before final
review. Each execution creates a separate candidate; never reuse old approval.
When machineStatus is publish-ready, run
take-a-repo review <outDir> --json and show its local URL. Let the user inspect
the resulting files and choose Approve or Request changes. Do not call the
review endpoint, click Approve, or write review.json on the user's behalf.
take-a-repo status <outDir> --json rechecks file bytes and returns feedback with
deliverable, source and claim references. Apply requested changes to those
inputs and recapture, then return for new user review.
Do not publish, upload, install globally, edit unrelated agent settings, or add background monitors as part of evidence collection. Even an approved file set does not authorize a new external destination. Report exactly what was run, what was verified, and what still needs user review.
Discovery depends on the host loading this skill or the consumer's AGENTS.md completion rule. Merely publishing an npm package does not make every LLM find or invoke it.