verifier-setup — scaffold this repo's /verify skill

Goal: leave the repo able to prove an engineering task works before it ships — run once, and it wires up everything the per-task /verify loop needs.

You are setting up — not verifying anything yourself right now. The /verify template lives at assets/verify.template.md (next to this skill). Parallels dev-local-setup (which generates a script + its skill doc): a setup skill that leaves behind reusable, repo-specific artifacts.

What this produces (the outputs)

Running verifier-setup end-to-end leaves the repo with:

1. A /verify skill — .claude/skills/verify/SKILL.md, the repo-tailored

per-task verification SOP (spawn a verifier sub-agent → drive the app → screenshot/video proof → open a PR with the proof embedded). Generated in Step 5.

1. A /dev-local skill + its script — scripts/dev-local.sh and

.claude/skills/dev-local/SKILL.md, via dev-local-setup (Step 2) if not already present. The one-command stack /verify depends on.

1. The driver skill installed — the playwright-cli skill for web apps (Step 2);

for non-web, the concrete exercise tool confirmed present.

Step 0 — Inventory what already exists (check before you add ANYTHING)

Before creating anything, take stock — the repo may already have some of this, under whatever name or layout its team chose. Look for the capability, not a specific filename; the paths below are only examples. For each, decide reuse as-is / adapt-extend / create fresh — never blindly overwrite working setup:

• A way to start the app — a one-command dev launcher, a Makefile/Procfile

target, docker-compose, package scripts (e.g. scripts/dev-local.sh, but any form counts).

• A prior verification SOP/skill — from an earlier run of this skill or the

team's own convention.

• A driver for the app's interface — a browser automation tool already available

(e.g. the playwright-cli skill), or the relevant API/CLI client.

• An existing test/e2e suite or checks — however organized.
• Sandbox/cloud-box config — anything giving isolated per-agent stacks.
• An evidence/artifact convention — where proof lands and how a reviewable

link gets published (a release, bucket, CI artifacts, etc.).

Every later step is conditional on this inventory: if a capability exists and works, reuse and adapt it (fill gaps, don't regenerate); only create what's missing.

Step 1 — Investigate the repo (don't guess)

Discover the real facts the generated skill will hardcode:

1. How the app is exercised — is it a web app (has a browser UI + a dev

server on a port), an API/service (HTTP endpoints, no UI), a CLI, or a desktop/mobile app? This picks the driver.

1. Stack launcher — is there already a way to start the app (any form — see

Step 0)? Note the up-command and the app URL/port. If none, Step 2 handles it.

1. Auth — is the primary flow login-gated? Is there a session/auth helper the

verifier can mint a session with (see e2e-setup)? Record it, or "n/a".

1. Regression checks — the repo's fast codified checks (type-check, lint,

unit, existing e2e commands) from package.json/Makefile/turbo.json/etc.

1. Proof upload — how a reviewable video URL is produced (a pr-evidence

GitHub prerelease via gh release upload is the default; a bucket/CI artifact works too).

Step 2 — Ensure the prerequisites exist (reuse-or-provide, per the Step 0 inventory)

For each, act on what Step 0 found — reuse if present, adapt if partial, create only if missing. Each check is idempotent; a no-op on what's already there:

• Dev stack. If any working way to start the app already exists (a launcher

script, Make/Procfile target, compose, package scripts), reuse it — read it for the up-command/port/services and move on (extend only if a needed service is missing). If there's none, scaffold one via dev-local-setup (don't hand-roll a launcher here). The generated /verify just needs a reliable one-command up.

• Driver skill.
• Web → install/confirm the playwright-cli skill (it documents + wraps

the browser driver). Ensure its binary is callable too (npx --yes @playwright/cli --version; install it + the chrome channel if missing). This closes the usual local gap where the browser driver was assumed but never installed.

• Non-web → confirm the concrete exercise tool exists (an HTTP client for an

API, the built binary for a CLI). No browser skill needed.

• Evidence dir. Ensure evidence/ is gitignored (proof output lands there).

Step 3 — Ask the user: local or sandbox?

Present the choice (default and recommend local — it's simpler to stand up):

• Local — one dev stack on the machine (scripts/dev-local.sh up). Best for a

single task at a time. Recommend this unless they need parallelism.

• Sandbox (crabbox) — an isolated cloud box per agent, for concurrent loops

or a fixed-port/single-instance stack. If chosen and not yet set up, scaffold via crabbox-setup; the generated skill drives the app in-box via cbx.sh pw.

Record the pick as the generated skill's default RUN_MODE (the other stays a documented fallback).

Step 4 — Confirm the driver

State the detected driver and confirm with the user (default playwright-cli for web apps). For non-web, name the concrete tool (e.g. curl/an HTTP script, the built CLI). This becomes DRIVER in the generated skill.

Step 5 — Generate .claude/skills/verify/SKILL.md

If a prior verification skill/SOP already exists (Step 0), update it in place — refresh the repo-specifics/placeholders, preserve any hand-edits the team added; don't clobber. Otherwise copy assets/verify.template.md → the repo's skills dir (.claude/skills/verify/ or the repo's convention) and fill every {{...}} placeholder from Steps 1–4: STACK_UP, APP_URL, RUN_MODE(+RUN_MODE_NOTE), DRIVER(+DRIVER_INSTRUCTION), AUTH_HELPER(+AUTH_INSTRUCTION), EXERCISE, REGRESSION_CMDS, EVIDENCE_UPLOAD, DATE. Delete branches that don't apply (e.g. drop the browser/video language for a non-web repo). Keep it to one screen — it's an SOP the agent follows, not an essay.

Step 6 — Hand off

Commit the generated skill (+ any .gitignore/dev-local changes). Tell the user:

• how to run it — "/verify before opening a PR (or when asked to verify), on a

branch with changes committed";

• the run mode chosen and how to switch;
• any prerequisite they must install/start once (e.g. Docker for infra, gh auth

for evidence upload).

Principles

• Check before you create; adapt, never clobber. Every artifact — dev-local,

the driver skill, e2e, evidence gitignore, the verify skill itself — is inventoried first (Step 0) and reused/extended when it already exists. Only create what's genuinely missing; preserve working setup and team hand-edits.

• Discover, don't assume. Stack command, port, auth, and checks come from the

repo — the generated skill hardcodes real facts, not conventions.

• Provision before you generate. The driver, launcher, and evidence/ exist

before the /verify skill ships, so it never fails on a missing tool.

• Right-sized. A web monorepo gets the full browser+video SOP; a CLI tool gets

a stdout-assertion SOP. Match the template to the repo; cut what doesn't apply.

• The output is a skill, not a run. verifier-setup scaffolds; /verify runs.