Content
81%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 well-engineered instruction skill: the phased workflow with verification gates, feedback loops, and rollback criteria is about as clear as workflow guidance gets, and templates make the deliverables concrete. The main improvable areas are trimming the diagram/section duplication and giving the exact Quint CLI syntax for invariant/witness selection at the gates.
Suggestions
Show the exact `quint run` invocation for selecting an invariant or witness (the actual flag and argument form) in the Phase 3 tool table and Phase 4 sections, so the verification gates are copy-paste executable.
Remove the ASCII phase diagram or the redundant prose re-descriptions of Phases 0–4 — keeping one of the two would cut ~30–40 lines and tighten the token budget.
Consolidate the repeated context-compaction advice (target 40–60%, 'under 300 lines', the status block) into a single section, and consider moving the detailed spec-change protocol or tool reference into a one-level-deep reference file to keep SKILL.md a leaner overview.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is overwhelmingly directive — 'Never modify the spec to make a failing verification pass', 'Run `quint typecheck` and fix all reported errors' — with no concept explanations Claude already knows. However, the 38-line ASCII workflow diagram duplicates the detailed Phase 0–4 sections that follow, and the compaction/compact-context advice ('Catalog the codebase compactly', 'Target: under 300 lines', the status block) recurs across three sections. This is 'minor instances that could be trimmed' (4), not fully lean (5) but well above noticeably verbose (2–3). | 4 / 5 |
Actionability | Highly concrete: exact commands ('quint typecheck', 'quint run', 'quint verify', '--max-samples 5000'), a copy-paste plan-format template, a research prompt template, a compaction status template, and a tool-usage table. The gap that keeps it at 4 rather than 5 is that the pivotal gate syntax is never shown — how to select an invariant or witness in `quint run` (e.g. the actual flag) is implied ('with witness as invariant', 'with those invariants') but not executable as written, and 'quint REPL session + quint REPL eval' is not a verifiable command form. | 4 / 5 |
Workflow Clarity | Exemplary sequencing: Phases 0–4 with a verification gate after every implementation step, rollback criteria in the plan template, explicit feedback loops ('If the gate fails: stop, diagnose, return to Phase 2', 'If invariant `[name]` fails after Step N, stop and return to planning'), a spec-change protocol requiring user approval, and failure-mode guidance for both violated invariants and unreachable witnesses. This matches the 5 anchor (explicit validation, error-recovery loops, checklist-style anti-pattern table). | 5 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/, or assets/), and the skill is a self-contained ~260-line workflow with well-organized, consistently headed sections; nothing is buried or nested. It stays at 4 rather than 5 because the single file carries material that could be split — the detailed per-phase bodies restate the overview diagram, and the tool-usage/spec-change detail could live in one-level-deep reference files to keep SKILL.md a leaner overview. This is 'good structure with minor organization gaps', not the ideal overview-plus-signaled-references split of 5. | 4 / 5 |
Total | 17 / 20 Passed |