Conciseness

The body is maximally telegraphic ('Reports are claims.', 'never pad', 'A silenced failure is a failure') and contains no explanation of concepts Claude already knows — every sentence is an operational rule, constraint, or template. It matches the 'lean and efficient; every token earns its place' anchor rather than the 4 anchor, which requires trimmable over-explanation, and none was found.

Actionability

Two complete copy-paste-ready prompt templates (specialist and generic modes) with fill-in rules, a verbatim phase-row table for the YOUR TASK field, a concrete delivery path convention (.tmp/scratch/reports/<YYYYMMDD>-<slug>.md), and exact adjudication rules ('"Tests pass" without the command output is not evidence; re-dispatch for the output'). This is fully executable guidance covering the common dispatch cases, matching the 5 anchor.

Workflow Clarity

The lifecycle is clearly sequenced (Decompose → Dispatch → Adjudicate → Pattern expansion) with explicit validation checkpoints: evidence-backed adjudication, independent review required for every code change ('Every request that changed code gets an independent review, from a reviewer other than the writer'), conflict resolution via falsifiable claims, and a bounded re-dispatch loop ('The two-attempt re-dispatch cap is a terminal safeguard… At the cap, report BLOCKED with the named gap'). This matches the 5 anchor's explicit-validation-plus-feedback-loop profile; the destructive/batch cap does not apply since validation (verify/validate/test phases, adjudication) is present.

Progressive Disclosure

Good structure: the overview holds the core dispatch mechanics while sub-agent contract and fix-cycle guidance live in well-signaled, one-level-deep reference files (both verified to exist and non-nested). It falls short of the 5 anchor on two real gaps: the pointer to ../parallel-work/SKILL.md is broken (no parallel-work sibling exists), and references/harness-notes/claude-code.md is present in the bundle but never linked from the body, so its content is undiscoverable.

Specificity

The description lists multiple concrete actions — 'Decompose substantive work into phases, dispatch each phase to a sub-agent, and adjudicate what comes back' plus judging reports, re-dispatching phases, and generalizing a user-named pattern instance — giving comprehensive coverage of the skill's lifecycle. It matches the 'multiple specific concrete actions; comprehensive coverage' anchor, not the level below, which requires minor coverage gaps.

Completeness

Both questions are answered explicitly: the 'what' is concrete (decompose, dispatch, adjudicate) and the 'when' opens with 'Use whenever a request asks for…' followed by concrete trigger phrases, plus an explicit exclusion ('Does not apply when your own prompt begins "Your ROLE_TYPE is sub-agent."'). This matches the 5 anchor exactly and is clearly above the 4 anchor, where the 'when' would be less specific.

Trigger Term Quality

Natural trigger terms users would actually say are comprehensively covered with synonyms: 'implementation, investigation, a fix, a review, or any change to files', 'read source', 'run a diagnostic', plus in-flight triggers like 'a report from a sub-agent needs judging' and 'a user names one instance of a pattern'. No common variants are missing, so it fits the 5 anchor rather than the 4 anchor's 'a few natural terms missing'.

Distinctiveness Conflict Risk

The delegation niche itself is distinct and the exclusion clause for sub-agent prompts helps, but the trigger surface is deliberately near-universal — 'Use whenever a request asks for implementation, investigation, a fix, a review, or any change to files — including small ones' — so it will fire alongside virtually every execution skill (code-review, security-review, run, etc.). It sits between the 'very broad; high overlap risk' (2) and 'minor overlap risk' (4) anchors, and the overlap is with many skills, not just closely related ones, so 3 fits best.