Content
70%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 highly actionable, well-sequenced orchestration skill with excellent validation gates and feedback loops, undermined by significant token bloat from repeated policy text and heavy inlining of content that belongs in reference files. Workflow and actionability are exemplary; conciseness and file structure are the clear weak points.
Suggestions
State the style-ref forwarding rule (writer-side phases get it, reviewer/auditor phases never do) once in a single table near the top and reference it from Phases 1, 2, 2b, 3, and 5 instead of repeating the full instruction at each phase.
Replace the two near-identical ~15-line helper-resolver bash blocks (style helper and audit verifier) with one canonical snippet cited via the existing integration-contract §2 link, or move both into a shared reference file.
Extract the Phase 1.5 reviewer prompt (including its long SCOPE LIMITS section) and the complete Phase 6.1 final-report template into reference files (e.g., references/contract-reviewer-prompt.md, references/final-report-template.md) to cut the body's inline bulk.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~946-line body repeats the style-ref forwarding rule in six places (header note, Phases 1, 2, 2b, 3, 5), restates the assurance-level derivation rule three times (Phases 0, 5.9, 6.0), and pastes two near-identical ~15-line resolver bash blocks. It is dense operational content rather than tutorials of concepts Claude already knows, so it sits above anchor 1 but clearly below the 'mostly efficient' anchor 3. | 2 / 5 |
Actionability | Guidance is copy-paste ready throughout: exact sub-skill invocations ('/paper-plan "$ARGUMENTS"'), complete resolver and detector bash scripts with exit-code handling (rg/jq numeric-claim detection, style-helper and audit-verifier resolution chains), a fully specified reviewer prompt, and complete checkpoint/report templates. The few conditional blocks (Phase 4.5/4.7 'if paper contains...') wrap concrete commands, so this does not fall to anchor 4. | 5 / 5 |
Workflow Clarity | Phases 0 through 6.1 are explicitly sequenced with checkpoints between each, a verbatim submission pre-flight checklist, verifier exit codes as the source of truth, per-verdict remediation (MISSING/STALE/BLOCKING_VERDICT/SCHEMA_INVALID), and feedback loops (max-3-round adversarial contract negotiation, 2 improvement rounds, re-validate after fixes). This matches the anchor-5 example pattern of sequence + explicit validation + error-recovery loops. | 5 / 5 |
Progressive Disclosure | The body is well-sectioned and signals one-level references clearly ('../shared-references/reviewer-backends.md', 'integration-contract.md', 'output-versioning.md'), but ~946 lines inline substantial content that belongs in separate files: the full Phase 1.5 reviewer prompt with its SCOPE LIMITS section, two duplicated resolver scripts, and the complete final-report template. No references/, scripts/, or assets/ bundle files exist, and the referenced ../shared-references/* paths do not resolve within this bundle, so this sits at anchor 3 rather than 4. | 3 / 5 |
Total | 15 / 20 Passed |