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 and well-sequenced deep-audit workflow with exemplary validation gates and feedback loops. Its weaknesses are token efficiency (heavily duplicated reviewer-protocol rules) and structure (reference-grade content inlined while all cross-file links point outside the bundle and are unresolvable as shipped).
Suggestions
Deduplicate the review-independence and reviewer-tier rules — currently restated in the opening blockquote, Key Rules, Phase 3, and the Submission Artifact section — into a single 'Reviewer protocol' section.
Move the 20-category issue taxonomy and the side-condition checklists to a references/ file (e.g. references/issue-taxonomy.md), keeping only a one-line summary per group in SKILL.md.
Normalize reference paths (some use ../shared-references/, others shared-references/) and either ship the referenced files with the skill bundle or remove the links.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly dense, non-obvious operational content, but reviewer-independence and model-tier rules ("same-family/provisional", "never use send_input", "always ultra") are repeated across the opening blockquote, Key Rules, Phase 3, and the Submission section, and Phase 5.5 embeds a padded shell fallback-resolution chain. Not 4: the duplication is substantive enough that the body could be tightened by roughly a quarter without losing information. | 3 / 5 |
Actionability | Provides exact copy-paste-ready artifacts: full spawn_agent prompt blocks with model/effort parameters, pdflatex compile commands, the complete PROOF_AUDIT.json schema, structured per-issue output formats, and a concrete python3 add_claim invocation. Not 4: the common cases (review prompt, fix record, compile check, artifact emission) are all fully executable with no gaps. | 5 / 5 |
Workflow Clarity | Phases 0–5.5 are explicitly sequenced with an objective acceptance gate, per-fix compile checks, re-review loops with MAX_REVIEW_ROUNDS, blind second review that re-enters Phase 2 on failure, a regression audit, and an honest unrecoverable-proof protocol. Not 4: validation checkpoints and error-recovery feedback loops are explicit at every stage, matching the anchor-5 example. | 5 / 5 |
Progressive Disclosure | The body is well-sectioned with clearly signaled links, but the ~100-line issue taxonomy, severity system, and side-condition checklists are fully inlined, and every referenced file (../shared-references/reviewer-backends.md, shared-references/reviewer-routing.md, ../shared-references/fan-out-pattern.md, ../lean-formalize/SKILL.md, shared-references/assurance-contract.md) lives outside the bundle with inconsistent path styles — none ship alongside SKILL.md. Not 4: the dangling, inconsistently-pathed external references and the heavy inline reference-grade content are more than minor organization gaps; not 2 because the section structure is real and navigation within the file is easy. | 3 / 5 |
Total | 16 / 20 Passed |