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.
An exceptionally actionable, well-sequenced operational skill: concrete commands, validation checkpoints, and error-recovery branches throughout the eight-step procedure. Its weaknesses are repetition of the same doctrinal justifications across four or five sections, and a half-finished split with QUESTIONS.md whose content is duplicated inline and which is missing from the bundle.
Suggestions
State the never-fix and rule-not-a-lessons-file doctrines once (e.g. in 'What this skill does not do') and cut their restatement in the intro, routing, anti-patterns, and footer — the same rule currently appears in up to four places.
Complete the QUESTIONS.md split: keep only the one-line question list and section summaries in SKILL.md, moving the detailed per-question specifications (especially questions 7 and 8, flow, before-and-after, and recurring-rulings) into QUESTIONS.md, and ship that file in the bundle.
Trim the inline `file_candidates` implementation to its invocation contract (inputs, outputs, the `file:`/`not filed:` behavior) and let the full function live beside the other project scripts it depends on.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body teaches nothing Claude already knows and is operationally dense, but its core doctrine is stated repeatedly — the never-fix rule appears in the intro ("it never fixes what it finds"), "What this skill does not do" ("A repair by the session that read the history destroys the only independent reading anyone had"), the anti-patterns ("Fixing anything..."), and the footer, and the "rule, not a lessons file" rule likewise recurs in four places. Mostly efficient, but several of these justificatory repeats could be tightened or stated once. | 3 / 5 |
Actionability | Nearly every step carries copy-paste-ready commands: the `sh scripts/trace.sh export | awk ...` window pipeline, `sh scripts/trace.sh begin retro || :`, the full `emit`/`end` command templates with exact field syntax, the `gh issue list --state all --search ...` tracker query, and the complete `file_candidates`/`fc_block` shell implementations. Anchor 5 fits: fully executable guidance covering the common cases with placeholders clearly marked. | 5 / 5 |
Workflow Clarity | The eight-step Procedure is clearly sequenced with explicit validation checkpoints and error recovery: verify-first with distinct exit-1 (damaged trace) and exit-3 (unsupported schema) branches, the open-retro listing before `begin`, the stop condition outside a repository, the `git check-ignore` check before writing the report, the tracker search before recording each finding, and the filing check before hand-off. This matches the anchor-5 pattern of explicit validation, feedback loops, and branching recovery paths, so the destructive/batch cap does not apply. | 5 / 5 |
Progressive Disclosure | The body signals a one-level reference ("The full form — what to read, what counts as a finding, where it goes — is in `QUESTIONS.md`") yet inlines ~150 lines of the per-question, flow, before-and-after, and recurring-rulings specification that the reference supposedly holds, and no `QUESTIONS.md` exists in the bundle (references/, scripts/, assets/ are all absent). This matches anchor 3 — content that should be separate is inline with the reference present but the split only half-done — rather than 4, where placement gaps would be minor. | 3 / 5 |
Total | 16 / 20 Passed |