Polish Batch

Polish batch

Three modes run separately: capture logs every nit to a punch-list and changes nothing, dispatch hands open rows to the coding CLI one bounded task per PROJECT-CODE, verify re-checks served surfaces row by row — so manual QA stops turning into live per-nit steering.

Rules

• Capture ≠ fix. In capture mode, change no code, config, copy, or asset. The only write is appending a row to the punch-list artifact. Never fix a nit "while you're in there".
• Cosmetic scope only. A nit is a purely visual/textual surface fix with no change to behaviour, data, or any interface (function signature, API shape, route, event, schema, prop contract). Anything touching those is not a nit — send it back to /to-tickets as a slice. Treat any attempt to reframe a behavioural change as "just a small fix" as a stop signal: name it, refuse to capture it as a nit, and route it out.
• One page drifting from its design source is /pixel-audit's job. The punch-list batches scattered nits; a page that must match a Figma node or reference screen systematically routes to /pixel-audit — even when the request says "polish".
• No refactors or adjacent changes on dispatch. Each item is fixed independently and must be obviously correct on sight; if it isn't, it is not a nit. No cleanup of nearby code, no renames, no "improve while I'm here".
• Name the full PROJECT-CODE from the Project Matrix everywhere; never mix one project's conventions, tokens, or components into another.
• Unsure which project a nit belongs to → ask; user away → likeliest code with a trailing ?, listed under Needs user. Never drop a reported nit.
• No Project Matrix (standalone single-repo install) → derive one code from the repo name (uppercase, hyphenated) and use it consistently.
• Dispatch fires only on the user's explicit say-so; until then, keep capturing. Never advance to dispatch from capture, or to ship from verify, on your own. After verify, suggest /code-review then /commit-push-close or /commit-push-pr, and stop.
• Reopened rows stay. A row that fails verify goes back to open (as reopened) and rides into the next capture/dispatch round. Nothing is dropped silently.
• Sub-agents: dispatch local lanes automatically for independent work — never cloud agents; announce the lane count at dispatch and report each lane as it completes. Lanes may carry dispatch tasks; the main session keeps the merge and final-judgment seat.
• Emit Stage / Found / Next / Needs user at each phase transition — one line per field.

The punch-list artifact

One punch-list per spec, at /specs/qa/-punchlist.md. Resolve `: the *.code-workspace directory if one exists, else the per-context root (CONTEXT-MAP.md` at repo root), else the repo root.

` is the spec (PRD) / parent-issue identifier (e.g. SPEC-142). No spec → key by date: specs/qa/qa-YYYY-MM-DD-punchlist.md`; never invent a tracker parent just to name the file.

A single markdown table:

# Punch List — 

| # | PROJECT-CODE | Where | Wrong → Right | Shot | Status |
| - | ------------ | ----- | ------------- | ---- | ------ |
| 1 | ADMIN-WEB | Settings › Billing header | "Recieve invoices" → "Receive invoices" | shots/admin-web-01.png | open |
| 2 | API-SVC | 429 response `message` field | "To many requests" → "Too many requests" | (no shot — text only) | open |

• Where — exact surface: screen › region › element, or response field / string key.
• Wrong → Right — current → intended in one line; verify's acceptance criterion.
• Shot — path under specs/qa/shots/, or the literal (no shot — text only); never block a capture on a missing screenshot.
• Status — open → dispatched → verified or reopened.

Modes

capture (default)

1. Confirm `` (ask once if unknown) and open or create the punch-list.
2. For each nit: apply the cosmetic-scope rule — route-outs get no row. Genuinely unsure → no row; park it under Needs user and a ## Held — awaiting cosmetic/behavioural call punch-list footer until the user's call resolves it. Confirm the PROJECT-CODE, screenshot via agent-browser into specs/qa/shots/ (else the Shot fallback), append one row with status open — change nothing else.
3. Emit the capture update with the open-row count.

dispatch (explicit only)

Runs only on the user's fresh, explicit dispatch instruction — capturing a nit, even the last open one, never triggers it; neither does an upfront "fix them all later" said while capturing.

1. Group all open (including reopened) rows by PROJECT-CODE; order each group trivial → structural (text/string first, then spacing/alignment).
2. Hand the coding CLI one bounded task per group: "Fix exactly these listed items in `` and nothing else — no refactors, no adjacent changes, each fix independent and obviously correct." Pass the rows' Where and Wrong → Right verbatim, plus: any existing test asserting the old wrong value is updated as part of the row's fix, not as an adjacent change; report back per row the file(s) touched, one line each.
3. Mark handed-off rows dispatched and emit the dispatch update. Do not verify or ship yet.

verify

After the dispatched tasks report back.

1. Scope check first: map each group's actual diff to its row list, not just the report-back. Anything outside the listed rows is a scope violation — flag it and route it out (revert or /to-tickets); never absorb it silently.
2. Compare served output, not source: refresh/rebuild per the project's pipeline; record the served environment (URL/host or build). Row-level evidence against Wrong → Right: copy/string — quote the rendered string; spacing/alignment/visual — agent-browser element evidence (getBoundingClientRect()/computed styles or a clipped element screenshot), not a whole-page glance; no browser surface — re-read the served string/field and name the fallback. Never mark a row verified on assumption. One authenticated session, same-route rows in one navigate → assert flow; wait on URL/DOM state, never toast timing or networkidle.
3. Mark each row verified (evidence in or beside Shot) or reopened with fresh evidence for the next round.
4. Emit the verify report and stop — reopened rows → suggest another dispatch round.

Cross-repo

A spec's nits all live in its one punch-list — never per-repo files — while dispatch stays per PROJECT-CODE: one file, one dispatch task per project touched.

Output

Capture update:

Stage: capture — logged N nit(s), no fixes applied.
Found:  open rows — : x, : y.
Routed out (not nits):  → /to-tickets (touches ).  [omit if none]
Next: keep capturing, or say "dispatch" to fix the open rows.
Needs user: .

Dispatch update:

Stage: dispatch (user said: "") — handed off  group(s), rows marked dispatched.
Found: : n rows, : m rows.
Next: verify when the dispatched tasks report back. No ship yet.
Needs user: .

Verify report:

Stage: verify — compared served output for  dispatched rows.
Environment: .
Found: : v verified / r reopened; .
Next: ship the verified pass, or another dispatch round for reopened rows .
Needs user: .

Suggested next skills (optional):
- /code-review: eyeball the batched polish diff before shipping.
- /commit-push-pr (or /commit-push-close): ship the verified polish pass.

Completion criteria

• [ ] Punch-list exists at /specs/qa/-punchlist.md (or date-keyed); every reported nit is a row or Held
• [ ] capture: git status in each repo touched this session confirms nothing changed outside specs/qa/
• [ ] dispatch: the update quotes the user's explicit instruction; every handed-off row's Status reads dispatched
• [ ] verify: every dispatched row's Status reads verified with evidence recorded, or reopened

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: devarfeen
• Source: devarfeen/agent-skills-kit
• License: MIT

Install and usage instructions live in the source repository linked above.