Prompt file imported from joefrancisGA/ArchLucid (
.cursor/commands/al-polish.md). Copyright stays with the author.
Polish batch (/al-polish)
One invocation, two models. The parent agent does not invent the suggestions and does not implement them.
| Phase | Model | Output |
|---|---|---|
| 1 — Suggest | Grok 4.7 High (grok-4.7-high) |
A short, code-grounded batch |
| 2 — Implement | GPT-5.6 Luna (gpt-5.6-luna-medium) |
Those items only, plus focused tests |
This is a local Task handoff, the same shape as /al-ui-rate. Do not launch it through /al-api. The Cloud Agent launcher cannot select these two models.
Arguments
/al-polish
/al-polish <count>
/al-polish "<focus>"
/al-polish <count> "<focus>"
/al-polish --suggest-only<count>— total suggestions. Default 20: 10 output and 10 usability. Minimum 1. Maximum 20. A higher number is clamped to 20."<focus>"— optional surface or job, such asdiagram question answersorsponsor export.--suggest-only— stop after Phase 1. No code changes.
With no count, ask for 10 output suggestions and 10 usability suggestions. With a count, split the batch in half. When that count is odd, the extra item is an output suggestion.
Allowlist
grok-4.7-high and gpt-5.6-luna-medium are outside .cursor/rules/Model-Allowlist-Override.mdc.
Before Phase 1, confirm this conversation already contains ok or yes authorizing both slugs for this /al-polish run. If it does not, ask once and wait. Name both display names and slugs, say why this command needs them, and state that the default without approval is Composer 2.5 slow (composer-2.5). Do not start Phase 1 in the same turn as that question. Do not substitute Composer or Grok 4.6.
What a suggestion is
A suggestion changes one existing surface so a reviewer cannot mistake an omitted field, a guess, or a missing link for a stored fact.
Output — the diagram, finding, path, question answer, or export a reviewer would forward.
Usability — the label, empty cell, or next action on that same path.
The first-review path is collect → diagram → findings → one question → one export. Prefer that path. Leave admin, catalog, webhook, and health-directory em dashes alone unless the focus text names that surface.
Guardrails
- Do not hide review workspace tabs.
- Do not add a score, a percentage, or a new metric.
- Do not invent a relationship, an owner, a path, or a resource.
- Do not call an Azure write API.
- Do not open a new product surface.
- Luna does not commit or push. After Phase 2, the parent opens one pull request against
masterfor the files Luna changed.--suggest-onlydoes not open a pull request. - Follow
.cursor/rules/Agent-Working-Tree-Safety.mdcbefore Luna edits a tracked file. A blocked dirty path is skipped, not overwritten. - Follow
.cursor/rules/shell-hygiene.mdcand.cursor/rules/shell-heartbeat.mdcfor every shell.
Workflow
Step 0 — Parse
- Read the count, the focus text, and
--suggest-only. - When no count is given, use 10 output and 10 usability. Otherwise clamp the total to 1–20, split it in half, and give any odd extra to output. Say both numbers before Phase 1.
- Complete the allowlist check above.
Phase 1 — Suggest (Grok 4.7 High)
Launch one Task subagent:
| Setting | Value |
|---|---|
subagent_type |
generalPurpose |
model |
grok-4.7-high |
run_in_background |
false |
description |
Polish suggestions (Grok 4.7) |
The prompt must include the count, the split, the focus text, the first-review preference, and the guardrails. Tell Grok to read the current code and skip a suggestion whose sentence is already on the surface.
Each suggestion uses this shape:
| Field | Content |
|---|---|
| Id | P-01 upward |
| Kind | output or usability |
| Surface | The screen or artifact |
| Current | The sentence or blank a reviewer sees now |
| Change | The sentence or behavior to ship |
| Files | Existing paths |
| Acceptance | What a test or the screen shows when done |
Print the batch in chat before Phase 2. Stop here when --suggest-only was set or the batch is empty.
Phase 2 — Implement (GPT-5.6 Luna)
Launch one Task subagent:
| Setting | Value |
|---|---|
subagent_type |
generalPurpose |
model |
gpt-5.6-luna-medium |
run_in_background |
false |
description |
Implement polish batch (Luna) |
Pass the batch verbatim. Luna implements every item that stays inside the guardrails. An item that needs a new API, a new relationship, or a dirty blocked path is skipped with a one-line reason.
Luna updates or adds the focused test for each shipped item and runs those tests. One scoped UI typecheck is allowed when the edits are in archlucid-ui. Luna does not commit.
Step 3 — Report
The parent checks git status --short and git diff --stat against Luna's file list. Then report:
| Field | Value |
|---|---|
| Suggested | Count from Phase 1 |
| Shipped | Count Luna implemented |
| Skipped | Id and reason |
| Tests | Command and result |
| Diff | Paths Luna changed |
Step 4 — Pull request against master
When at least one item shipped, open a pull request. Skip this step when --suggest-only was set, the batch is empty, or Luna shipped nothing.
- Fetch
origin/master. - Ask the user to reply with the exact branch name
cursor/al-polish-<short-topic>, and wait for that reply before creating it fromorigin/masteror running any branch-switch, commit, or push command. Do not commit onmaster. - Stage only the files Luna changed for this batch. Leave unrelated dirty and untracked files unstaged, including SQL, package zips, and
next-env.d.tswhen those were already dirty and were not in Luna's file list. - Commit only Luna's paths with
git commit --only -- <Luna-file-list>so any unrelated paths staged before this run remain outside the commit. The message says why a reviewer was seeing a guessed or blank value. - Push with an explicit refspec:
git push -u origin HEAD:cursor/al-polish-<short-topic>. Do not push that commit tomaster. Do not force-push. - Open the pull request with base
masterand that branch as the head.
Add the pull request URL to the report.
