AGENT SKILL

red-pen

Turn a raw design red-pen — a pile of screenshots plus dictated critique of the generated scenarios — into deduped, correctly-routed, harvested changes to the AXP design harness. Use when the user dumps a critique of the scenario builds and wants it captured and actioned without losing anything, over-harvesting, or trusting the wrong source. The harvest half of the design-quality loop; pairs with product-loop (the generate half).

License 许可
Apache-2.0
所属仓库
torkelo/scenes2
最近更新
2026-11-14
Stars
★ 0
Forks
0

01 前言 Front Matter

name
red-pen
metadata
{ "source": ".hatch/_skills/red-pen/SKILL.md", "generated": "hatch@v1.2.0" }
description
Turn a raw design red-pen — a pile of screenshots plus dictated critique of the generated scenarios — into deduped, correctly-routed, harvested changes to the AXP design harness. Use when the user dumps a critique of the scenario builds and wants it captured and actioned without losing anything, over-harvesting, or trusting the wrong source. The harvest half of the design-quality loop; pairs with product-loop (the generate half).

03 SKILL.md 正文

red-pen

The intake for a human critique of the generated scenarios, and the counterpart to /product-loop: product-loop generates and judges; red-pen turns the critique into routed changes to the harness in product/. Read product/README.md for the homes and invariants first. Run this pipeline in order; do not jump to editing.

1. Parse the dump faithfully

The input is voice dictation (filler, run-ons, transcription errors — "insect" → "inset", profanity as emphasis) plus screenshots that reference specific scenario/vN. Parse the intended meaning, map every screenshot to its scenario and version, and when a point is genuinely ambiguous ask rather than guess.

2. Verify — trust neither the dictation nor the builders

Ground each claim against the actual render (screenshot the story on :6006 — never spawn a server, see no-spawning-storybook — or read the attached crop). Both the critique and the builders' own caughtFixed reports have been wrong. Mark what you could not verify (behind a tab, scroll behaviour) rather than asserting it.

3. Dedupe to recurring patterns — the highest-value step

Collapse the notes into a small set of patterns grouped by underlying cause, not by scenario. The recurrence is the insight (inconsistency is the meta-defect): "eyes zigzag / nothing lines up / on top of itself" is one row-alignment finding, not fifteen.

4. For each pattern, ask the higher-level question

Do not just log the point-fix. For each cluster, name the foundational change that would stop the whole class ("what do we author so a hard black border is never reached for") and only then the specific fix beneath it. Generalisation first.

5. Route each finding to exactly one home

Per product/README.md: a composition rule → the matching docs/design/* reference; a component-agnostic fundamental → docs/design/foundations.md; a non-checkable judgment → docs/design/taste.md; a shipped value bug → component-specs/COMPONENT-FIXES.md; a cross-cutting state/behaviour → component-specs/CONTRACTS.md; a component to build → a spec + BUILD-QUEUE.md. Never restate a captured value (one home). Separate genuinely-new findings from confirmations of existing rules (confirmations add priority evidence, not new entries).

6. Flag tensions — never silently overwrite a rule

When a finding conflicts with an existing rule, or is a judgment call, surface it as a decision for the user rather than editing over it. Do not over-harvest: a scenario-specific one-off that is already an instance of a general rule gets no entry of its own.

7. Output, then execute

First a review artifact — (a) recurring patterns, (b) new rules by home, (c) confirmations, (d) tensions needing a decision, (e) where you'd push back. After the user's steer (or straight away for the unambiguous ones), apply the edits, Prettier-clean them, and report where each finding landed. Finally, note which new rules the next product-loop round should be able to test — closing the loop back to generation.

Guardrails

Dedupe before routing · one home per value · verify, don't assume · generalise over point-fix · flag, don't overwrite · preserve the user's voice and intent · don't over-harvest.

04 同仓库更多 Skills