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 exemplar of actionable, well-sequenced instruction: concrete commands, a hard verification gate with a fix-recheck loop, and specific error recovery. Its weaknesses are token efficiency — the same invariants restated in four places and long digressions into host-system behavior — and a complete absence of progressive disclosure, with everything inlined in one 550-line file.
Suggestions
Move the auto-link/MCP receipt semantics (Phase 7 note) and the facts_backstop/chronicle extraction lifecycle (Phase 5) into a one-level-deep reference file (e.g. references/extraction-behavior.md) and link to it from the relevant phases, cutting the top-level file to the core pipeline.
State each invariant once: the "not ingested until the checklist passes" rule and the facts_backstop rationale each appear in four places (Contract, phases, Verify gate, Output Format, Anti-Patterns) — keep the HARD GATE statement and cross-reference it elsewhere instead of restating it.
Convert Phase 7's ~25-line auto-link prose into a short bullet list of rules (who gets attended edges, when to run `gbrain extract --stale`, what to do over MCP vs stdio) to preserve the content at a fraction of the tokens.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | There is no generic padding — no explaining what meetings or diarization are — but key invariants repeat across sections: "A meeting is NOT fully ingested until" appears in the Contract, the Verify gate, Output Format, and Anti-Patterns; the facts_backstop lifecycle is explained in the sidecar section, Phase 5, "The loop", and Output Format; and Phase 7's auto-link/MCP note runs ~25 lines on host-system behavior. This fits anchor 3 (mostly efficient but could be tightened) rather than 4, because the repetition and digressions exceed "minor instances of over-explanation". | 3 / 5 |
Actionability | Copy-paste-ready commands throughout — `gbrain whois <email> --json`, `gbrain timeline-add {person-slug} {date} "Attended {meeting-title}"`, `gbrain day {date}`, `gbrain sync --source <id> --no-pull`, `gbrain extract --stale` — plus a complete normalized-record YAML schema, a fill-in meeting page template, exact dedup criteria ("date ± 1 day + attendee overlap ≥ 50%"), and named receipt fields to check. This matches the 5 anchor: fully executable guidance with specific examples covering the common cases. | 5 / 5 |
Workflow Clarity | Nine numbered phases in strict sequence, a "HARD GATE" verification checklist (V1–V6) with deterministic checks, an explicit feedback loop ("The loop: fix → re-check → fix, until every item passes"), and an error-recovery section ("`put_page` returns `revision_conflict`... re-read and merge; never overwrite a person page from an old read"). This is exactly the 5 anchor: explicit validation steps, feedback loops for error recovery, and a checklist for a complex process. | 5 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/, or assets/), so all ~550 lines live in SKILL.md. Cross-references to other skills are clearly signaled ("skills/enrich/SKILL.md", "docs/guides/life-chronicle.md", "../conventions/brain-first.md#identity-exact-match"), but reference-grade material — the auto-link/MCP receipt semantics and the facts_backstop/chronicle extraction lifecycle — is inlined rather than split into one-level-deep files. This is anchor 3 (content that should be separate is inline), not 2 because structure and signaling are good, and not 4 because essentially nothing is offloaded to progressively disclosed files. | 3 / 5 |
Total | 16 / 20 Passed |