Skip to content

SKILL.md

Latest commit

 

History

History
88 lines (72 loc) · 5.29 KB

SKILL.md

File metadata and controls

88 lines (72 loc) · 5.29 KB
Raw
name care-review
description Full review of a CARE frontend (care_fe) diff via two parallel lens agents — (1) what problem the change solves + whether the code makes it legible, and (2) whether the approach is the simplest one or overengineered — condensed into one report. Use for "review/critique my changes", "review the diff", "full review", "review against develop / the last commit". Add "thorough" for a heavier cross-model ensemble. Defaults to diffing against develop; suggests rather than edits.
user-invocable true
argument-hint [develop | commit | working | <file>] [thorough]

CARE Review (orchestrator)

Two lenses, run as parallel subagents and condensed into one report. Each agent reconstructs intent independently — you can't judge a solution until you know the problem — and the orchestrator reconciles their readings:

  1. Intent & legibility — what does the change do, what requirement does it fulfill, does the code convey that on its own? → care-diff-review
  2. Approach — given that problem, is this the simplest solution or overengineered; what can be simplified or reused? → care-technical-review

Flow

  1. Resolve the diff once → a temp file. Default is everything the branch changes vs develop:

    git diff $(git merge-base develop HEAD) > /tmp/care_review.diff && wc -l /tmp/care_review.diff

    Overrides only when asked: last commit (git show HEAD), working only (git diff + git diff --staged), or a named file. Always write to the file and read that — never rely on inline terminal output, which truncates on large diffs and wastes turns re-running git. List the changed files. Do not read the commit message, PR body, or branch name — intent must come from the code.

  2. Pick the mode:

    • Default — lens-split. Two agents in parallel, both on Claude Opus (see Models).
    • Thorough — ensemble. Triggered by a thorough/deep argument or an explicit ask for a heavier/more-careful review. Run the lens-split across two models (Opus + GPT-5.5) for cross-provider diversity.

  3. Dispatch the agents in parallel. Give each agent the diff-file path and instruct it to: read /tmp/care_review.diff (scope is already resolved — do not re-resolve or re-run git), reconstruct intent independently from the code (no commit message), apply its lens, and return its findings to you — it must not confirm with the user (you own the single confirm).

    • Default: Agent A → care-diff-review; Agent B → care-technical-review. Both on Opus.
    • Thorough: run that A+B pair twice — once on Opus, once on GPT-5.5 (up to 4 agents in parallel).

  4. Condense into ONE report — reconcile, don't concatenate:

    • Bottom line — one or two sentences: is it sound / mergeable? For a refactor-only diff, the behavior-preserving yes/no.
    • Intent — one or two sentences (what + the requirement), with confidence. Reconcile the agents' independent intent readings: they agree → high confidence; they diverge → flag the ambiguity (the code isn't conveying its intent).
    • Worth deciding — only the 1–2 findings that actually matter: correctness, a regression in the other usages of a shared component/hook/util the diff touched, or real overengineering. This is the signal; keep it short.
    • Optional / FYI — everything else (nits, style, micro-simplifications), clearly demoted so it doesn't dilute the decision.
    • Out of scope — one line for anything deliberately left untouched.

    In thorough mode, note where the two models agreed vs. diverged — agreement across families is high confidence; divergence is a look-closer signal. A clean diff can legitimately return "intent clear, approach proportionate, nothing to change." Don't manufacture findings.

  5. Confirm with the user. Lead with the reconstructed intent — "is this the problem you were solving?" — since everything hinges on it. A mismatch means illegible code or a wrong solution. Then walk the (few) suggestions. Don't edit until approved.

Models

Concrete picks — overridable at invocation; update this block when availability changes.

Mode Agents
Default (lens-split) both lenses on Claude Opus
Thorough (ensemble) full lens-split on Claude Opus and on GPT-5.5, reconciled

Dispatch caveats (verify in your host):

  • Claude Code's Agent tool spawns Claude models only — the GPT side of the ensemble needs a host that exposes GPT (e.g. VS Code Copilot).
  • If the host can't set a subagent's model, run /care-review thorough as two passes — once with the session on Opus, once on GPT-5.5 — then do the reconcile pass yourself.
  • If the host won't spawn subagents at all, fall back to a single combined pass: read both lens rubrics and do one review on the current model. Still valid — you just lose parallelism and cross-model diversity.

Working agreement (inherited by every agent)

Suggest first, don't edit. Smallest possible diff — no scope creep, no adjacent "improvements". Don't change a component's contract or invent logic to paper over a problem. Less code, not more. New code reads like the file around it (CLAUDE.md).