AGENT SKILL

architecture-change-reviewer

Use before multi-file changes (>3 files) or cross-module work — emits a 5-line verdict (bounded-context-guardian / input-contract-tests / error-handling-enforcer / complexity-anti-drift / minimal-change-verifier) and blocks on no / unclear. Triggers on multi-file change, new module, cross-module refactor, plan before implementation.

License 许可
MIT
最近更新
2026-10-28
Stars
★ 2
Forks
0

01 前言 Front Matter

name
architecture-change-reviewer
bucket
engineering
description
Use before multi-file changes (>3 files) or cross-module work — emits a 5-line verdict (bounded-context-guardian / input-contract-tests / error-handling-enforcer / complexity-anti-drift / minimal-change-verifier) and blocks on no / unclear. Triggers on multi-file change, new module, cross-module refactor, plan before implementation.
disable-model-invocation
false

02 资源文件 2

03 SKILL.md 正文

Architecture Change Reviewer

Overview

Multi-file changes are reviewed before code lands, by mapping each affected Core Skill to a one-line verdict. If any verdict is "no" or "unclear", the plan is incomplete — fix it, don't ship it. Cross-skill conflicts (e.g. "split the bounded context" vs "one task") resolve by sequencing, not merging. Tracer-bullet tickets and the [decision] / [implementation] tags live in writing-plans, not here. ACR stops at the all-yes verdict.

When to use

  • Planning a change that touches 3+ files
  • Adding a new module / subsystem / bounded context
  • Refactoring across module boundaries
  • A task where 2 or more Core Skills plausibly apply
  • Before implementation begins, on a change with cross-file blast radius

When not to use

  • Single-file typo fix
  • Single-file bugfix with a clear single Core Skill in scope
  • Pure documentation or comment changes
  • Trivial rename inside one file
  • Decomposing a feature into ordered vertical-slice tickets — that is writing-plans (one outcome per tracer bullet, one tag).

Procedure

  1. Enumerate the planned files. Stop if fewer than 3.
  2. For each Core Skill, write a one-line verdict:
    • bounded-context-guardian: are module boundaries respected?
    • input-contract-tests: is the change covered by tests for the 5 boundary classes?
    • error-handling-enforcer: is every failure path typed, non-empty, EXIT-documented?
    • complexity-anti-drift: does the plan's declared structure keep one abstraction level per function/module (no god-function / god-file intent, no planned deep nesting or duplication)? Smell thresholds guide, line counts never decide.
    • minimal-change-verifier: is this one task? Will the diff stay in that scope?
  3. If any verdict is "no" or "unclear" — stop, go back to the responsible Core Skill and fix the plan.
  4. If 2 Core Skills conflict (e.g. bounded context says split, minimal change says one task) — split the work into separate tasks. Don't merge scopes.
  5. If no Core Skill applies — the change is too small to need this reviewer. Proceed without.
  6. Hand off. writing-plans consumes the all-yes verdict block and writes the ordered tracer-bullet task list to plans/<feature>.md (see writing-plans/references/handoff-acr.md).

Rationalization Table

Excuse Reality
"Just 4 files, I can eyeball the verdict" 4 files × 5 Core Skills = 20 checks. Eyeballing skips 18.
"It's obviously a small change, skip the review" "Obviously small" is the rationalization that ships shotgun surgery. Enumerate anyway.
"Two Core Skills conflict, just pick one" Conflict means both apply. Sequence them, don't drop one.
"Plan is in my head, no need to write it down" Head-plan ≠ reviewable plan. The 5-line verdict block lives in plans/<feature>.md (SSOT).
"Cross-skill review is bureaucratic overhead" Cross-skill review catches the diff that slips past every individual Core Skill.
"Just decompose tasks, skip the verdict" Task decomposition is writing-plans' job. ACR without verdicts is no gate.

Red Flags - Stop and Start Over

  • Plan description does not list affected files
  • Any of the 5 Core Skills has no written verdict
  • Verdict uses "probably" / "should" / "looks good" instead of yes/no/unclear
  • 2+ Core Skills in conflict and the plan does not sequence them
  • Implementation has begun before the 5-line verdict is committed to plans/<feature>.md
  • "I'll review during PR" — review is pre-implementation, not post-hoc

Required Baseline

Zero tolerance: no multi-file change proceeds without all 5 Core Skill verdicts written as yes/no/unclear. Any "no" or "unclear" verdict blocks implementation. Conflicts are sequenced, never merged.

Acceptance Criteria

  • [ ] Plan has >= 3 affected files
  • [ ] Each Core Skill has a written verdict (one line)
  • [ ] All 5 verdicts are "yes", "N/A with reason", or trigger "no"/"unclear" → block. "no" or "unclear" = mandatory block, no override. (no "probably", no "should")
  • [ ] No Core Skill conflict unresolved
  • [ ] Implementation does not begin before this review is satisfied
  • [ ] Architecture review verdict is committed in plans/<feature>.md (the SSOT)
  • [ ] complexity-anti-drift verdict is grounded in declared structure, not in line-count estimates of code that does not exist yet

Verification

  • grep -c "bounded-context-guardian:" plans/<feature>.md ≥ 1 (verdict block present)
  • grep -c "minimal-change-verifier:" plans/<feature>.md ≥ 1 (all 5 verdicts in plans/<feature>.md)
  • diff <(grep -E "^\s*affects:" plans/<feature>.md | awk '{print $2}' | sort -u) <(git diff --name-only | sort -u) produces no output (planned file list == actual diff scope, no scope creep)
  • Each of the 5 verdict lines is present in plans/<feature>.md

Example

See references/example-verdict-block.md for a complete 5-verdict example showing how to use this skill to review a multi-file change before implementation. See references/dimension-semantics.md for the pre-impl vs post-impl split of each dimension.

04 同仓库更多 Skills

winter6205
winter6205/arthurpower

writing-plans

Use when turning a buildable spec into ordered tracer bullets after architecture-change-reviewer. Covers vertical slicing, dependency order, and headroom for the implementer. Triggers on 写计划, 拆任务, 任务分解, 制定方案, 跨文件改动, 计划.

★ 2
winter6205
winter6205/arthurpower

wayfinder

Use when a chunk of work is more than one agent session can hold — plan it as a shared map of decision tickets on your issue tracker, and resolve them one at a time until the way to the destination is clear.

★ 2
winter6205
winter6205/arthurpower

verification-before-completion

This skill should be used when claiming "完成" / "done" / "fixed" / "passing" / "ready to ship", before trusting a subagent success report, or when checking "对照计划" / "对照 spec" — whether this run's delivery matches its basis.

★ 2
winter6205
winter6205/arthurpower

using-agent-skills

Use when starting a new session, switching tasks, or unsure which arthurpower skill applies. Routes incoming work to the matching skill via flowchart before invoking. Activates on phrases like "which skill should I use", "skill check", "pre-task skill check", "how to start", "router", "skill routing", or when multiple competing skills could apply. Does not auto-invoke user-invoked authoring skills (skill-authoring, self-evolving-rules).

★ 2
winter6205
winter6205/arthurpower

test-driven-development

Use when implementing logic, fixing a bug, or changing behavior — write the failing test first (RED), then minimal code (GREEN), then refactor. Triggers on "write the failing test first", "Prove-It", or a bug that needs a reproduction test before the fix.

★ 2
winter6205
winter6205/arthurpower

systematic-debugging

Use when debugging a test failure, build break, unexpected runtime error, flaky test, or hanging process. Triggers on phrases like "fix this bug", "getting an error", "failing test", "flaky", "hanging", "halfway through debugging", "this used to work", "intermittent failure", "stack trace shows X", or any case where behavior does not match expectations. Prevents shotgun debugging.

★ 2