Content
78%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.
A dense, expert-grade meta-skill whose structure and file routing are exemplary: every reference is real, one level deep, and loaded exactly when needed, and validation is explicit and risk-tiered. The weaknesses are stylistic and size-related — hyper-compressed multi-clause sentences that violate the skill's own plain-sentences rule, and a body that exceeds the 8,000-byte budget the skill itself states — plus the absence of any inline worked example to ground its central decision rules.
Suggestions
Apply the skill's own 'one idea per sentence' rule to its densest sections — e.g. split 'Sediment first' into its three actual steps (search provenance, keep-with-citation when found, apply admission when not) — since clauses fused with dashes cost re-reads even when they save bytes.
Trim the body under the 8,000-byte SKILL.md budget it cites by consolidating the overlapping length rules ('Lean prompts for current strong models', 'Sol-first portability', 'Portable length control preserves content') or moving one to a reference named at its point of use.
Inline one compact worked example of the central move — a case list restated as the condition it proxies — so 'Conditions, not cases' is executable without first opening a reference file.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | No padding and no explanation of concepts Claude already knows — every rule carries content, so it clears anchor 4 ('efficient; minor instances that could be trimmed') rather than anchor 3. It is not a 5 because the brevity frequently fails the 'preserves clarity' test: sentences fuse two or three rules with dropped connectives ('Provenance decides how hard to look, not what stays: search for a test that asserts the line… Provenance found → … None found → apply admission (…)'), the exact pattern the skill's own 'Plain sentences' rule labels a defect, and the ~10 KB body sits above the 8,000-byte SKILL.md budget it itself cites. | 4 / 5 |
Actionability | Concrete anchors are present and executable: `bun run test:skill-guards`, `bun test`, `tests/codex-skill-prompt-budget.test.ts`, `references/evaluate.md`, a mode table routing each request type to a specific reference file, and a fully specified completion-report shape. It is not a 5 because the core rules ('Sediment first', 'Match freedom to fragility', 'name the condition X is a proxy for') are decision standards with no inline worked example — an example rewrite, finding, or report lives only in the references — and not a 3 because the commands, paths, and per-mode done conditions cover the common execution path. | 4 / 5 |
Workflow Clarity | A coherent sequence is present (read the standard section the mode names → apply the always-hold rules → pick mode → read its reference → validate → report) with an explicit validation checkpoint ('After any edit, `bun run test:skill-guards` runs the mechanical guards in about a second'; behavior changes get a targeted eval 'or an explicit skip reason in the report') and per-mode done conditions. It is not a 5 because the top-level flow is implied by section order rather than stated, and error-recovery detail (what to do when validation fails) is delegated to the references rather than closed with a feedback loop at this layer. | 4 / 5 |
Progressive Disclosure | All referenced bundle files exist (references/new-skill.md, edit-skill.md, review-skill.md, respond-to-review.md, evaluate.md — verified on disk) and each is signaled at its point of use: the Modes table routes 'You are X → Read `references/<mode>.md` → Done when', and `references/evaluate.md` is named inside the validation rule that needs it. References are exactly one level deep, the split between always-loaded cross-mode rules and mode-specific procedure is appropriate, and navigation is easy — a clear match to anchor 5. | 5 / 5 |
Total | 17 / 20 Passed |