Conciseness

The body assumes Claude's competence throughout — no padding about what CLIs or reviews are — and every section delivers operational rules ("Execute all workflow sections in order", "Always pass `--cwd` with the absolute project root path"). Not 5 because pointers like "Full contract: `second-opinion --help`" recur across three sections and several sentences are so clause-dense (the 10-line Error Handling paragraph) that trimming would improve them; not 3-2 because there is no known-concept explanation and virtually no filler.

Actionability

The invocation is concrete and executable — ".agents/skills/second-opinion/scripts/second-opinion <mode> [options]" plus a filled-in quick-mode example with real flags ("quick \"your question here\" --cwd /path --foreground") and unambiguous pass/omit rules for `--target`, `--timeout`, `--cwd`, and `--foreground`. Not 5 because the option surface, exit-handling contract, and identity rules are deferred to "second-opinion --help" rather than shown, leaving gaps Claude must fetch at runtime; not 3 because the given commands are copy-paste executable and the flag rules cover the common cases.

Workflow Clarity

Execution is clearly sequenced: mode table → execution rules → wait-command handling ("Execute the exact printed wait command ... until terminal"), with explicit validation checkpoints in exit codes, the stderr JSON error object ("{\"error\": ...}\"), the CLI-failure record taxonomy, and recovery feedback ("For a timeout, suggest a larger `--timeout` or a narrower `--range`"). Not 5 because the per-mode step sequences live in the workflow files and the recovery rules are scattered through dense prose rather than a checklist; not 3 because validation signals and feedback loops are explicit and thorough.

Progressive Disclosure

Structure and signaling are good — a workflow table with clear links, and "Lane resolution, merge rules ... [references/multi-lane.md]" is a well-signaled one-level-deep reference that exists in the bundle. But four of the five linked workflow files (workflows/review.md, challenge.md, audit.md, quick.md) do not exist in the bundle, breaking navigation to the core per-mode instructions, and the dense Error Handling and Configuration walls are inlined where a reference file would fit. Not 4 because dangling primary references are more than a minor organization gap; not 2 because references are clearly signaled, not buried, and multi-lane.md is properly split out.

Specificity

"Load for a cross-model review, challenge, audit, or quick consult" names the domain and enumerates the four action modes, but each action is a single unelaborated word ("review", "challenge") and the mechanism (running an external AI CLI) is absent. Not 4 because coverage has gaps (the detect mode is omitted and no action is described beyond its name); not 2 because several concrete mode names are given.

Completeness

The "what" is stated (cross-model review/challenge/audit/consult) and "Load for a..." is explicit trigger guidance answering "when". Not 5 because the "what" is thin — the external-CLI mechanism and outputs are unstated in the description; not 3 because the when-clause is explicit rather than weakly implied.

Trigger Term Quality

"review", "challenge", "audit", and "quick consult" are natural phrases users would say when needing this skill, giving good keyword coverage. Not 5 because common synonyms are missing — "second opinion" itself, "cross-check", and model-name triggers (Claude/Codex) that the summary field uses do not appear; not 3 because the four terms present are the primary user phrasings.

Distinctiveness Conflict Risk

The "cross-model" qualifier is the distinguishing marker, but bare "review" and "audit" overlap heavily with code-review and security-review skills, so a request like "review my code" could route here wrongly. Not 4 because the generic action words carry substantial overlap risk with closely related skills; not 2 because the cross-model framing does carve out a specific niche.