Conciseness

The body is two lines — an HTML generation comment and "Load and follow the shared skill at `~/.agents/skills/ci-fix/SKILL.md`" — with zero padding and no explanation of things Claude already knows. Not 4: there is literally nothing that could be trimmed; every token earns its place.

Actionability

The single directive is concrete (a specific path and a specific action), but all task-executable guidance — how to enumerate failures, apply fixes, commit, push, watch runs — is delegated to the referenced file, so the body itself is incomplete guidance for the task. Not 4: an executable pointer is not executable task guidance; the body contains no steps, commands, or examples of its own.

Workflow Clarity

The delegation action itself is unambiguous, but the body exposes no sequence and no validation checkpoints for the multi-step fix→commit→push→watch→repeat loop it fronts; the simple-skill exception doesn't apply because the underlying task is a complex, risky (push-triggering) workflow whose checkpoints are invisible here. Not 1: the one step written is coherent and unambiguous rather than missing.

Progressive Disclosure

One clearly-signaled, one-level-deep reference — exactly the right shape for a wrapper — and no inlined content that belongs elsewhere. Not 5: the pointer targets an absolute machine-local path (`~/.agents/skills/ci-fix/SKILL.md`) outside this bundle (no references/ directory exists, and the path does not resolve in this environment), so navigation depends on environment setup rather than bundled files.

Specificity

Enumerates many concrete actions — "enumerates failing workflows/jobs/steps", "determines the root cause", "proposes a fix plan, applies code changes, commits, pushes, watches the new runs, and repeats until checks pass" — giving comprehensive, concrete coverage of the workflow. Not 4: there are no gaps in coverage; the full pipeline from diagnosis to green checks is explicitly listed.

Completeness

Explicitly answers both what (the full investigate→fix→push→re-run loop) and when ("Use when the user says 'fix CI'..."), with concrete trigger phrases. Not 4: the 'when' is fully explicit with quoted user utterances and input forms, plus negative scoping ("Do NOT use for authoring new workflows, or for read-only diagnosis with no fix intent — use gha-log-reader for that").

Trigger Term Quality

Covers natural user phrasings comprehensively: "fix CI", "CI is red", "checks are failing", "the build is broken on my PR", "fix the failing jobs", plus the input pattern "pastes a failing CI run URL". Not 4: it includes multiple synonym phrasings and an input-format trigger, not just a few good keywords.

Distinctiveness Conflict Risk

Clear niche (fixing failing CI) with explicit disambiguation against adjacent skills — read-only diagnosis is routed to "gha-log-reader" and workflow authoring is excluded — minimizing wrong-skill triggering. Not 4: the negative triggers don't just reduce overlap risk, they name the competing skill.