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.
The body is an exceptionally rigorous, fully executable verification manual with a clear failure-triage workflow and built-in validation loops — near the ceiling on actionability and workflow clarity. Its weaknesses are volume and monolithic structure: historical findings, timing measurements, and world inventories are inlined where reference files would keep the acting core lean.
Suggestions
Move the "Findings" section (per-world invalid-merge counts, M0.6 named merges, sync-commit history) and the scratch-break table under "What a pass proves" into a references/findings.md, keeping only the failure-signature samples inline that the triage decision tree actually uses.
Split the world/command/mutant inventories in "What it checks" into a references/worlds.md so SKILL.md keeps the property statement, the rule taxonomy, and the run/triage workflow at the top level.
Trim time-sensitive operational numbers (CPU seconds, load-average anecdotes, exact leaf pins) into the baseline file's documentation or a short 'pins' reference, so the acting instructions stay valid as the counts drift.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The prose is information-dense with no padding or explanation of concepts Claude already knows, but the body is a ~45KB monolith whose historical sections ("Findings" counts from the sync commit, scratch-break tables, CPU-second measurements, pinned counts like 813,356 leaves) carry time-sensitive operational record beyond what the acting workflow needs. It fits anchor 3 — mostly efficient but could be tightened — rather than 4, since substantial material could be trimmed or moved without losing actionability. | 3 / 5 |
Actionability | Guidance is fully executable: exact commands with flags ("pnpm check:effect-ledger --all --world=<shard>", "--case='<identity or part of one>'", "--record-baseline"), concrete failure identities to paste back (e.g. "internal blocks/pp | accepted retype 1 | own enter 1 mid"), and a worked reading of each example. This matches the copy-paste-ready, common-cases-covered level-5 anchor. | 5 / 5 |
Workflow Clarity | The workflow is clearly sequenced with pervasive validation and feedback loops: run a lane, print a single case with --case, classify the failure under the explicit six-way decision tree (broken merge / ledger hole / modeling error / law failure / changed pin / intended change), fix, re-run, then re-record the baseline. The coverage-record section adds a literal three-item checklist before moving the sync commit, matching the level-5 anchor including for this batch/destructive context. | 5 / 5 |
Progressive Disclosure | The body has strong internal structure (numbered rule taxonomy, tables, headers) and points to real repo artifacts (effect-ledger-classes.json, the scripts, the plan doc) instead of duplicating them, but there is no bundle at all — findings history, world definitions, and the coverage record are all inlined in one file that clearly could be split into reference files. This lands on anchor 3 (content that should be separate is inline) rather than 2, because the structure and navigation within the file are genuinely good. | 3 / 5 |
Total | 16 / 20 Passed |