Content
77%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body is exceptionally actionable and workflow-safe: exact commands, port tables, dual-check PR pre-flight, and validated checklists for destructive operations. Its weaknesses are length and offloading — tool-internal scanner/reaper semantics are inlined in a 330-line single file with a dangling reference instead of being split into one-level-deep reference files.
Suggestions
Move the reaper/scanner/ledger internals (the 出生来源与 PR 关联 prose and the trim retry semantics) into a one-level-deep reference file, keeping only the owner-facing rules in SKILL.md.
Replace or repair the '*(internal reference removed)*' stub with a real reference path or delete the sentence, since a dangling pointer misleads navigation.
Trim defensive edge-case narration (EXDEV behavior, multi-round retry accounting) to the one-line rule plus a pointer, cutting the body closer to a lean overview.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Most content is non-obvious project-specific state (port tables, Redis boundaries, finish/retain semantics) that Claude cannot know, and commands are given without padding, so it is not score 2. But sections like the 出生来源与 PR 关联 scanner paragraphs ('API 每五分钟扫描...不伪造原 finish 的完成回执') and the trim/reaper retry semantics describe tool-internal behavior an executing agent does not need, matching 'mostly efficient but includes some unnecessary explanation or could be tightened'. | 3 / 5 |
Actionability | Every operation ships copy-paste-ready commands: the worktree:new invocation with flags, .env heredocs with exact values, the port-offset derivation table, finish/retain/check invocations, and diagnostic commands like 'pnpm check:worktree-port-offset'. Placeholders like {feature-name} are standard templates, and edge cases (restore, sandbox policy, stash footer) each get an exact command or literal text to write. | 5 / 5 |
Workflow Clarity | The lifecycle is explicitly sequenced (recall → main sync check with ahead/behind computation → in-flight PR two-step check → create → env → verify → finish) with explicit validation checkpoints throughout: Redis isolation verification, the 安全核查 checklists before creation and cleanup, --check receipt verification for destructive finish, and error-recovery guidance (what to do when out of sync, when archive/delete is interrupted). Destructive removal is guarded by owner/protection/process checks, satisfying the feedback-loop requirement. | 5 / 5 |
Progressive Disclosure | Headers are clear and the document is navigable, but it is a ~330-line monolith with no bundle files: content that clearly belongs in a separate reference (reaper internals, scanner/ledger behavior, port-offset deep design) is inlined, and one pointer is a dead stub ('详细设计见 *(internal reference removed)*'). This matches 'some structure but content that should be separate is inline'; it is not score 2 because sections are well-labeled rather than a structureless wall. | 3 / 5 |
Total | 16 / 20 Passed |