Content
66%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 exceptionally strong on workflow discipline — sequenced checklists, explicit validation guards, and named error-recovery paths — and its guidance is concrete and largely executable, backed by a complete, trigger-labeled reference bundle. Its weakness is token efficiency: dense multi-clause prose, heavy restatement across sections, and duplication of reference content inline make a 756-line core that could be substantially tightened without losing any rule.
Suggestions
Cut the Quick Reference table or reduce it to non-duplicative pointers: its rows restate the restatement check, the id-derivation rule, the sibling check, and the two-guard rule almost verbatim from the body, doubling the token cost of the same policies.
Move the 20-line id-derivation bash snippet and the full frontmatter field spec into references/observation-log.md (which already documents them), keeping in the core only the new-observation.sh invocation and the one-line fallback pointer — the snippet is the single largest inline block duplicating reference content.
Split multi-clause sentences and drop the inline 'why' asides (e.g., the parenthetical justifications after each guard in the Session Start Protocol); each rationale already lives in the referenced section it names, so the core can state the rule and the pointer alone.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The 756-line body is noticeably verbose for a companion/logging skill: rules are restated across sections (the restatement check, checkpoint rules, and archival policy each appear in prose, in the snippet comments, and again in the Quick Reference table), and single sentences routinely run 60+ words with nested parentheticals and 'why' asides (e.g., the workspace-anchoring paragraph). It is not 1 because the text never explains concepts Claude already knows — every sentence carries skill-specific policy — but it far exceeds 'some unnecessary explanation that could be tightened'. | 2 / 5 |
Actionability | Guidance is mostly executable: a complete copy-paste id-derivation bash snippet with guards (noclobber create, collision check, listing guard), exact script invocations ("bash scripts/new-observation.sh --body <file> <slug> \"[ABSOLUTE PATH]\"", "scripts/set-fields.sh \"[ABSOLUTE PATH]\" <plan>", "scripts/diff-staged.sh <staged> <live>"), and a full frontmatter template enumerating every field. It falls short of 5 because some procedures give only the outcome vocabulary ("ERROR is not reconciled... DIFFERS is classified") or defer entirely to a reference section without the inline command, leaving minor gaps. | 4 / 5 |
Workflow Clarity | The Session Start Protocol is a seven-item tick-box checklist with explicit ordering, and validation is pervasive with named feedback loops: the storage probe is 'split before any branch is chosen', 'an empty scan in a log known to be non-empty is a broken command until proven otherwise', failed writes get bounded retry rules, and the write snippet guards every failure mode it names. This matches the top anchor — clear sequence, explicit validation steps, error-recovery loops, and a checklist for a complex process. | 5 / 5 |
Progressive Disclosure | Scored against the actual bundle: all nine referenced files (weekly-review, skill-authoring, observation-log, observation-scan, observation-families, signals, environments, migration, starter-principles) and all named scripts exist beside SKILL.md, and each pointer carries an explicit load trigger ('Load when a review triggers', 'Load only when the Session Start Protocol detects a legacy log'), one level deep. It is not 5 because substantial detail that has a reference home is also inlined in the core — the 20-line id snippet and the full frontmatter spec duplicate observation-log.md material, and the Quick Reference table re-states body content. | 4 / 5 |
Total | 15 / 20 Passed |