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.
A strong, well-sequenced instruction skill: the five-phase pipeline is concrete and executable, and destructive operations are unusually well-guarded with validation gates, reversibility, and explicit error recovery. Its main costs are repetition (reversibility and test-before-bulk restated multiple times) and a monolithic inline Phase 5 rubric that inflates token cost without adding navigation value.
Suggestions
Move the Phase 5 cull rubric (6-axis tables, verdict thresholds, grounding labels, cluster budgets) into a reference file (e.g., references/cull-rubric.md) and keep a short summary plus the merge mechanics in SKILL.md, cutting the body roughly in half.
Deduplicate repeated assertions: state reversibility and test-before-bulk once each (plus one Anti-Patterns entry) instead of restating them across sections, and drop or compress the 'Contract'/'Output Format' conformance-test boilerplate.
Provide at least one concrete prompt template for the LLM distinctness judge and the T1/T2 synthesis pass so those steps are copy-paste executable like the gbrain commands.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with skill-specific judgment criteria Claude could not know (6-axis cull rubric, grounding labels, cluster budgets), but it repeats itself: reversibility is restated roughly five times, test-before-bulk three times, and the distinctness-guard 'hard veto' appears both in the merge section and again in Anti-Patterns; the trailing 'Contract' and 'Output Format' sections exist only 'for the conformance test'. Not 4 — these are several padded passages, not minor trimmable instances. | 3 / 5 |
Actionability | Concrete executable commands throughout ('gbrain query "type:concept" --limit 10000 --json', get/put/delete/restore/history/revert, 'gbrain pricing set'), plus exact frontmatter schemas, numeric verdict thresholds (≥85% ELITE, <55% DELETE), and full output templates. Not 5 because the LLM-judge and synthesis passes are directed ('LLM: are these the same idea?') without copy-paste-ready prompts, though the refusal to hardcode a similarity threshold is explicitly justified. | 4 / 5 |
Workflow Clarity | Phases 1-5 are explicitly sequenced, and this destructive/batch skill has real validation checkpoints: the merge-quality gate that rejects half-merges, the logged distinctness-guard veto, spot-check count reduction, a 3-5-cluster test-before-bulk probe, three explicit undo paths, and per-error recovery (resume_command, revision_conflict re-read). Clear sequence with explicit feedback loops; the destructive-operation cap does not apply because validation is present. | 5 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are absent), and external references (../conventions/quality.md, test-before-bulk.md, agent-operator-protocol.md, _brain-filing-rules.md) are one level deep and clearly signaled in blockquotes at point of use, with well-organized section headers. Not 5 because the ~200-line Phase 5 cull rubric is inlined monolithically where a reference file would keep SKILL.md an overview, and the conformance-test boilerplate pads the tail. | 4 / 5 |
Total | 16 / 20 Passed |