Conciseness

The body is information-dense with no concept explanations Claude already knows, but several sections could be tightened or relocated: the ~75-line 'Tracked configuration coverage' triage with its table, the 'Control-plane note' path inventory, and long exception parentheticals like "(for example `using-git-worktrees` pointing at `/git:worktree`) or sits inside a workflow the user started explicitly (`backlog-batch-delivery` steps, `executing-plans` finishing with `branch-finish`)". This is more than the 'minor instances' of the 4 anchor, but well short of the padding of the 2 anchor.

Actionability

Guidance is copy-paste executable throughout: `python3 ~/scripts/harness-sessions.py <slugs> --since <N> --json`, `just audit-ignored-config .omp .pi`, `git log --oneline -- <hook script>`, `codex debug prompt-input`, a concrete YAML snippet for `agents/openai.yaml`, a worked example ("$optimize-harness claude hooks since:14"), and a complete output template. Common cases are covered by specific commands, matching the 5 anchor.

Workflow Clarity

The 'Execution shape' gives a clearly numbered sequence ('Run one deterministic scan first', fan out audits capped at 3 concurrent, structural transcript privacy boundary) with explicit validation gates ("A finding with no evidence is dropped, not softened"; "Never render `unknown/not observed` as `unused`"), error-recovery paths for invalid tokens ("stop and print the two valid token sets above. Do not guess") and absent harnesses ("reported as `not installed / not configured` and skipped"), plus a read-only guardrail ("Do not mutate config during the audit"). Sequence, checkpoints, and a dimension checklist are all present, satisfying the 5 anchor rather than the 4.

Progressive Disclosure

All three bundle references exist and are clearly signaled with when-to-read guidance (`references/harness-map.md`, `references/session-usage-evidence.md`, `references/agent-harness-portability.md`), and they are one level deep within the bundle. However, the long inline 'Tracked configuration coverage' section and the portability mapping rules (which duplicate `references/agent-harness-portability.md`) are content that could live in reference files, and reference files point onward to out-of-bundle docs (`docs/agent-session-transcripts.md`) — minor organization gaps that fit the 4 anchor rather than the clean split of the 5 anchor.

Specificity

The description enumerates seven concrete audit areas ("config efficiency, hook overhead, permission friction, plugin and MCP cost, skill metadata portability, dead config, and measured session friction") and names all seven target harnesses explicitly. Coverage of what the skill audits is comprehensive rather than leaving minor gaps, so it matches the 5 anchor over the 4 anchor.

Completeness

It explicitly answers both questions: what ("Audit one or all supported agent harnesses... for [seven named areas]") and when ("Use when invoked as `$optimize-harness [harness ...] [focus]`, or when the user asks to optimize or compare agent harness configuration") with concrete trigger phrases, matching the 5 anchor exactly.

Trigger Term Quality

Natural trigger phrases are present ("Use when... the user asks to optimize or compare agent harness configuration", plus terms like "audit", "hooks", "permissions", "plugins", "MCP" and the explicit "$optimize-harness" invocation). A few natural variations users might say (e.g. "speed up", "tune", "clean up config") are missing, which keeps it at 4 rather than 5; it is well above the 3 anchor's 'missing common variations' because the core intent words are covered.

Distinctiveness Conflict Risk

It occupies a clear niche — multi-harness configuration auditing with seven named harnesses — and its triggers include the explicit "$optimize-harness" invocation syntax, making conflict with other skills minimal. It is not the 4 anchor because no meaningful overlap risk with closely related skills is identifiable.