Content
63%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 dense with high-value, non-derivable operational detail and concrete commands, organized into a clear phased workflow. It is weakened by repetition of authorization rules, scattered version pins, and the absence of any bundle files, which forces protocol and policy detail to live inline in an over-long SKILL.md.
Suggestions
Split the memory-verbs protocol, visibility/readback rules, and budget-packing guidance into files under references/ and keep one-line pointers in SKILL.md; the two opening blockquotes alone would cut roughly 50 lines from the main file.
State the capture-authorization rule once (e.g., in the Contract section) and reference it from Phases 2 and 4 instead of restating it three times; move version requirements (gbrain >= 0.43, v0.18.0+) into a single requirements note.
Embed the readback/validation step directly in the Phase 2 write sequence (write -> recall to confirm -> only then report success), so the mutating write loop has an explicit validation checkpoint rather than one stated only in Output Format.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Most content is genuinely novel domain knowledge (CLI verbs, visibility rules, the shallow-ls trap) rather than concepts Claude already knows, but there is noticeable tightening to do: the authorization/suppression rules are restated three times (intro, Phase 2, Phase 4), the two opening blockquotes are dense walls of edge-case policy, and version pins (v0.43+, v0.18.0+, gbrain >= 0.43) are scattered inline rather than isolated in a deprecated/version section. Not a 2 because the bulk of the tokens carry non-derivable operational detail. | 3 / 5 |
Actionability | Concrete, copy-paste-ready commands throughout: `gbrain entity "<name>"`, `gbrain recall --query 'zebra telescope' --budget-tokens 75 --budget-policy query_first --json`, `find <dir> -name '*.md' | wc -l`, `gbrain sweep --once`, plus an explicit citation example and verb signatures like `remember(fact, provenance, ttl?)`. Not a 5: a few steps stay abstract ("Check backlinks: who references this entity?" and Phase 3's "Pull context") and the example commands lack surrounding invocation context for the common cases. | 4 / 5 |
Workflow Clarity | A clear phased sequence (Phase 1 through 4, with 1.5/2.5 interludes) with numbered steps and check-before-write loops (entity lookup before creating pages, verify counts against `gbrain sources list`, confirm writes only after readback). Not a 5: for a mutating, batch-writing skill the readback/validation checkpoints live in the Output Format section rather than being embedded in the Phase 2 write sequence, leaving the write loop itself without an inline validation step. | 4 / 5 |
Progressive Disclosure | Sections are clearly headed and external references are one level deep and signaled ("See `skills/conventions/brain-first.md`", "Contract: `docs/protocol/MEMORY_VERBS_v1.md`"). However, no bundle files exist (no references/, scripts/, or assets/ directories), so the detailed material those files would hold — the memory-verb protocol, visibility policy, budget-packing rules, and the ~220-line body itself — is inlined in SKILL.md instead of split out, and one reference ("Read `budget_packing`") has no path at all. This fits anchor 3: structure present, but content that should be separate is inline. | 3 / 5 |
Total | 14 / 20 Passed |