Content
87%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A tight, high-signal instruction skill: it adds only what Claude wouldn't already know, gives a concrete cleaning-pass decision tree, and cleanly delegates detail to sibling skills. The main gap is an explicit validation checkpoint in the cleaning-pass workflow before the PR is opened.
Suggestions
Add an explicit final verification step in the cleaning pass, e.g. 'For every file touched, record the verdict (keep/fix/dropped) and confirm none remains an unrecoverable question before opening the PR,' to give the workflow a concrete validation checkpoint.
Make the section ordering read as an enforced sequence rather than parallel checklists — e.g. note that the security skill's review runs before the cleaning pass (already implied) and state the order in which to walk Code, Security, and Docs so Claude doesn't skip gates.
For the PR-description guidance, add one short worked example (bug fix naming a diagnostic-ladder rung, feature naming audience + use case) so the 'permanent index' instruction is copy-paste ready rather than abstract.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Lean throughout — it assumes Claude's competence, stating only the extra OSS bar beyond CI/CLAUDE.md/naming and never explaining what a PR, secret, or semver is. Every section earns its tokens with non-obvious defaults ("push = publish," "incomplete is worse than absent"). | 3 / 3 |
Actionability | Gives concrete, executable tests — the literal stranger-question prompt for the cleaning pass, the helps/recoverable/unrecoverable decision tree, and named removals (tmp.ts, notes.md, snapshots). Directives like "Use `draft: true`, defer the PR, or just don't ship" are copy-paste ready for instruction-only work. | 3 / 3 |
Workflow Clarity | The cleaning pass is sequenced with a clear decision tree, but there is no explicit validation/verification checkpoint confirming each artifact's verdict before opening the PR, and the section ordering (Code, Security, Docs, Cleaning pass) reads as a checklist rather than an enforced sequence with feedback loops. | 2 / 3 |
Progressive Disclosure | Well-organized single-file overview that delegates depth via clearly signaled one-level-deep references to sibling skills (security, links, etiology, naming, code-ts, code-react, pedantic) and docs/AGENTS.md; no nested/deep reference chains and no bundle files to misroute. | 3 / 3 |
Total | 11 / 12 Passed |