Craft Review

A craft review of one diff through five lenses, each in an isolated sub-agent, aggregated without reranking across lenses.

• Domain-Driven Design — is this the right model?
• Code Smells — is the structure decaying? (Fowler, Refactoring 2nd ed., ch. 3)
• Clean Code — is it readable at the line? (Martin, Clean Code 2nd ed.)
• Pragmatic Programmer — is it built deliberately and kept flexible? (Hunt & Thomas, 20th-anniversary ed.)
• Philosophy of Software Design — is complexity leaking in? (Ousterhout, 2nd ed.)

Isolated contexts stop a clean bill on one lens from masking a failure on another, and stop the lenses' overlapping items (primitive obsession, Law of Demeter) from being deduped away.

Process

1. Pin the diff

Fixed point = whatever the user named (SHA, branch, tag, main, HEAD~3). Default when unspecified: everything not on main plus uncommitted work — git diff $(git merge-base HEAD main)...HEAD and git diff HEAD.

Capture the diff command(s) and git log <fixed-point>..HEAD --oneline once. Confirm the ref resolves and the diff is non-empty before spawning anything — a bad ref or empty diff fails here, not inside five sub-agents.

Within that diff, skip paths no lens has anything to say about: lockfiles (package-lock.json, yarn.lock, Cargo.lock, …), anything under a directory literally named generated/dist/build/vendor, and binary assets (images, fonts, audio). A repo's own docs (read in step 2) may name further repo-specific exclusions — e.g. a vendored third-party export living under docs/ — those take precedence over this generic list.

2. Gather shared context

Read the project's domain model if present — CONTEXT.md (or the CONTEXT-MAP.md contexts touching the changed files) and any docs/adr/ entries covering the changed area. Note any repo-documented coding standard (CLAUDE.md, AGENTS.md, CODING_STANDARDS.md, CONTRIBUTING.md, etc.) — including any review-scope note it gives (paths that aren't authored application code). Proceed silently if none exist.

This context is pasted to every sub-agent so a documented standard or ADR can override a lens heuristic.

3. Spawn five sub-agents in parallel

One message, five Agent calls, general-purpose for each. Resolve <skill-dir> to this skill folder's absolute path.

Every prompt includes:

• The diff command(s) and commit list from step 1.
• The shared context from step 2.
• The brief: "A documented repo standard or ADR overrides any heuristic in your checklist. Every finding is a judgement call — label it so. Skip anything tooling already enforces. For each finding: name the item, quote the offending hunk, give the fix in one line. Under 500 words."

Per lens, add:

• DDD sub-agent: "Read <skill-dir>/ddd.md. Check the diff against every item. Also flag drift from the CONTEXT.md ubiquitous language and any contradiction with an ADR."
• Smells sub-agent: "Read <skill-dir>/smells.md. Check the diff against every smell. Use the metric thresholds only as tie-breakers."
• Clean Code sub-agent: "Read <skill-dir>/clean-code.md. Check the diff against every rule."
• Pragmatic Programmer sub-agent: "Read <skill-dir>/pragmatic.md. Check the diff against every item."
• Philosophy of Software Design sub-agent: "Read <skill-dir>/philosophy.md. Check the diff against every red flag, then the principles behind them."

4. Aggregate

Present the five reports verbatim (or lightly cleaned) under ## Domain-Driven Design, ## Code Smells, ## Clean Code, ## Pragmatic Programmer, ## Philosophy of Software Design. Do not merge or rerank across lenses — a hunk flagged by two lenses is signal; report it under both.

When two lenses directly contradict — one says do X, another says do not-X on the same hunk — keep both findings but mark the losing one "overridden by ", following this precedence: DDD → Pragmatic Programmer → Clean Code → Code Smells → Philosophy of Software Design. This resolves head-to-head conflicts only; it is not licence to rerank non-conflicting findings by lens.

End with one line per lens: finding count and the worst finding within that lens. No single cross-lens winner — that reranking is what the separation exists to prevent.