Conciseness

Lean and dense throughout: a task-to-agent decision table, exact output-contract templates, and a one-line rule of thumb ("if you'd want the subagent's output in 1/3 the tokens, pick cavecrew"). Nothing explains concepts Claude already knows — even the "Why this exists" section is three sentences of skill-specific rationale (tool results injected verbatim into main context), not generic padding. Matches the 'every token earns its place' anchor.

Actionability

Highly actionable for an instruction-only skill: the decision table gives exact task→agent mappings, output contracts are specified verbatim ("<path:line-range> — <change ≤10 words>" / "totals: N🔴 N🟡 N🔵 N❓"), and chaining patterns are concrete ("Spawn 2-3 `cavecrew-investigator` calls in one message (different angles: defs vs callers vs tests)"). Not 5: no example of the actual prompt/invocation text to hand a cavecrew agent, a minor gap; not 3 because the guidance is specific and executable, not pseudocode-level.

Workflow Clarity

The main "Locate → fix → verify" chain is numbered 1-2-3 with the reviewer as an explicit audit step, and the builder contract embeds a verification checkpoint ("verified: <re-read OK | mismatch @ path:line>"). Not 5: there is no explicit recovery loop after the reviewer finds issues (the chain ends at "audits the diff") and the mismatch case has no stated retry step; not 3 because checkpoints are present and clearly sequenced, and this is not a destructive/batch skill so no cap applies.

Progressive Disclosure

Single-file skill with no references/, scripts/, or assets/ bundle and no external files referenced; nothing that belongs in a separate file is inlined — output contracts and chaining patterns are short and needed at a glance. Sections (when to use vs alternatives, output contracts, chaining patterns, what NOT to do) are well-organized and navigable, matching the well-organized single-file case.

Specificity

Names three concrete per-agent capabilities — "(locate code)", "(1-2 file edit)", "(diff review)" — plus the mechanism "Their output is compressed, so main context lasts longer", which matches the 'several specific actions, minor gaps' anchor. Not 5: it never spells out what "locate code" spans (finding definitions, callers, uses), so coverage is not fully comprehensive; not 3 because it lists three distinct concrete actions rather than only 1-2.

Completeness

Explicitly answers both questions: the opening "When to delegate to `cavecrew-investigator` ... or `cavecrew-reviewer` (diff review) instead of working inline or using `Explore`" is a full when-clause with per-task triggers, and the what is stated via each agent's function plus "Their output is compressed, so main context lasts longer". Voice is third person ("Their output"), so no voice penalty applies. Not 4 because the when-clause is already explicit and task-mapped, not merely present-but-vague.

Trigger Term Quality

Contains natural, sayable terms — "locate code", "1-2 file edit", "diff review", "delegate", "Explore", "context" — giving good coverage a user or Claude would echo. Not 5: common variations like "find where X is defined", "what calls Y", "search the codebase", or "review my changes/diff" are absent; not 3 because the terms go beyond bare domain naming and include the natural task verbs.

Distinctiveness Conflict Risk

The compressed-output delegation niche is distinct and the description explicitly contrasts itself with "working inline or using `Explore`", matching 'mostly distinct; minor overlap risk'. Not 5: the trigger domain (code location, edits, diff review) inherently overlaps with built-in Explore and code-review skills, so mis-trigger risk is more than minimal.