Content
73%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.
The protocol content is exceptionally actionable and the workflow is the best-case anchor: explicit state machine, mechanical gates, and layered recovery procedures. The clear weakness is conciseness — the body spends large token budgets on self-justifying meta-commentary (known-gap disclosures, historical rationale for field formats, repeated legacy-name callouts) that directly contradicts its own "context window = RAM" principle, and the spine carries tables that should live in references.
Suggestions
Move the self-justifying meta-disclosures (the unscoped-Bash known-gap note, the changelog step-field no-prefix rationale, the leash-override exception history) out of SKILL.md into a reference or project changelog — they explain design history, not runtime behavior, and cost resident tokens on every activation.
Relocate the File Lifecycle Matrix and File Ownership table bodies to references/file-formats.md (both are lookup material, not activation-time reading), keeping only the read-before-write rule and a pointer in the spine.
Collapse the repeated legacy `plan_` directory-name callouts (appearing in the plan-dir definition, bootstrap, recovery step 0, and Git Integration) into a single note in references/file-formats.md and reference it once.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~490-line body is noticeably verbose: it carries several padded meta-disclosures — the "Known gap — unscoped Bash on read-only agents" paragraph, the multi-paragraph justification for why the changelog `step` field carries no plan-id prefix ("Do not 'fix' the apparent inconsistency..."), the "Disclosed exception — the leash-override 'continue'" digression, and repeated legacy `plan_` naming callouts across at least four sections. None of this is explanation of concepts Claude already knows (so not a 1), but the skill ironically preaches "Context Window = RAM. Write to disk immediately" while inlining substantial rationale that belongs on disk. | 2 / 5 |
Actionability | Fully executable throughout: copy-paste-ready commands (`node <skill-path>/scripts/bootstrap.mjs new "goal"`, `--pre-step`, `reset-attempts`), exact gate semantics ("Exit code 2 emits one of four `GATE:FAIL` slugs — `[no-plan]`, `[wrong-state]`, `[leash-cap]`, `[iteration-cap]`"), a complete file tree, per-file lifecycle and ownership matrices, and a worked commit-tag derivation ("`plan-2026-07-14T051317-317362c4` → `[plan-2026-07-14-317362c4/iter-3/step-2] desc`"). `<skill-path>` resolution is itself specified. Anchor 5: specific examples cover the common cases. | 5 / 5 |
Workflow Clarity | Exemplary sequencing and validation: a Mermaid state machine plus a transition table with explicit triggers, a Pre-Step Gate that hard-blocks on failure, a "4-item Post-Step Gate", the REFLECT "3-phase gate: Gate-In (7 mandatory reads), Evaluate... then Gate-Out", and feedback loops for error recovery (revert-first leash with 2 attempts, 3-Strike Rule, a 13-item numbered recovery-from-context-loss procedure). Matches anchor 5 (explicit validation steps, feedback loops, checklists); not 4 because checkpoints are explicit at every transition, not merely most. | 5 / 5 |
Progressive Disclosure | Good bundle structure scored against the actual files: 9 real `references/*.md` files all exist and are annotated in the References section, per-state detail is correctly pushed to `scripts/modules/state-<state>.md` via the emit-state router (all 5 modules verified present), and `agents/ip-*.md` is explicitly framed as an optional install with a documented fallback rather than a broken path. Not 5 because the SKILL.md body itself still inlines substantial detail — the File Lifecycle Matrix, File Ownership table, and compression-protocol format blocks — that belongs one level deeper, leaving the "spine" dense rather than overview-shaped. | 4 / 5 |
Total | 16 / 20 Passed |