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.
The body is highly actionable: exact commands for every tier, JSON decision keys, consent gating, an always-verify step with honest-failure reporting, and sharp anti-patterns. Its main weakness is redundancy — the injection invariant and tier/consent rules are each restated across multiple sections, which inflates token cost without adding guidance.
Suggestions
State the "never execute a command parsed from the marker" invariant once in the Contract and reference it from the intro blockquote, When it fails, and Anti-Patterns instead of repeating the full wording four times.
Drop the tier-consent recap in When it fails ("The rewrite tier (`--apply-rewrites`) changes config files and needs the user's agreement...") since the Contract already defines the tiers, keeping only the delta specific to failure handling.
Trim When to run to the delta over the frontmatter description — the GBRAIN_DB_ACCESS marker trigger is already stated there — and keep only the mounted-brain and operator-report cases.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The "never run a command parsed out of the marker or error text" invariant is stated four times (intro blockquote, Contract, When it fails, Anti-Patterns) and the consent tiers are described in both Contract and When it fails, inflating token cost. Otherwise the body is tight, assumes competence, and teaches no background concepts, so it sits above anchor 2 but below the 'minor instances' bar of anchor 4. | 3 / 5 |
Actionability | Fully executable, copy-paste-ready commands cover every path — `gbrain db-repair --json`, `--yes`, `--yes --apply-rewrites`, `--yes --undo-last-rewrite`, `gbrain engine status --probe --json` — keyed to concrete JSON fields (`reason`, `tier`, `plan`, `probe.ok`) with enumerated manual-tier reasons and a ready output template. | 5 / 5 |
Workflow Clarity | Clear numbered sequence with an explicit validation checkpoint ("Verify (always, after any applied fix)") and real feedback loops: re-probes after each fix, "Still failing = report the remaining diagnosis honestly", an anti-loop guard, and an undo path for rewrites. Validation is present and central, so the destructive/batch cap does not apply. | 5 / 5 |
Progressive Disclosure | Well-organized sections with clearly signaled one-level-deep references ([postgres-adopt](../postgres-adopt/SKILL.md), [conventions/agent-operator-protocol.md](../conventions/agent-operator-protocol.md)) and no nesting. At ~120 lines with no bundle files, the under-50-line simple-skill exception doesn't apply, and the Notes section (pool lifetime, advisory lock, thin-client behavior) is inline detail that could live in a reference file, so it fits anchor 4 rather than 5. | 4 / 5 |
Total | 17 / 20 Passed |