| name | harness-audit |
| description | Periodic review of claudemax's own scaffolding against the current Opus capability. Flags load-bearing vs vestigial pieces so the harness stays sharp as models improve. Run after every major model release or quarterly, whichever comes first. |
/harness-audit — keep the scaffolding sharp
Models improve. Yesterday's helpful scaffolding becomes today's friction. This skill audits claudemax against the current model and surfaces what should be deleted, simplified, or kept.
What it checks
- Skill catalog overlap — re-runs the overlap audit against
docs/SKILL_CATALOG.md. Flags any skill whose purpose now overlaps another (e.g., a primitive the umbrella has subsumed).
- Prompt verbosity — sample the system prompts in
packages/runtime/src/prompts.ts. Flag any prompt > 1500 chars whose every clause is not load-bearing for the current Opus baseline.
- Router rules — re-evaluate
packages/core/src/router.ts heuristics against the current Opus default capabilities. Flag rules that no longer change behavior on the current model.
- Dark-patterns hook coverage — audit
.claude/hooks/ invocation against current-model failure modes. Flag hooks that haven't fired in N runs and modes that have started slipping through.
- Cost-guard thresholds — re-evaluate 70 / 90 / 95 % thresholds against actual run-cost distribution from
memory.runs.
- Skill description currency — every SKILL.md
description: should still reflect what the skill actually does. Flag drift.
Output
A HARNESS_AUDIT.md artifact in the repo root with sections:
## load-bearing (keep)
- <item> — <one-line reason it's still earning its place>
## redundant (delete or merge)
- <item> — <what overlaps it, why deletion is safe>
## simplify
- <item> — <what to cut from it without losing the load-bearing piece>
## measurement gaps
- <thing we cannot decide without data> — <what to instrument>
The output is human-reviewed; this skill does NOT auto-delete anything.
When to run
- After every major Anthropic model release (
/harness-audit --against opus-X.Y).
- Quarterly as hygiene.
- When
cmax doctor flags credit consumption climbing without proportional effectiveness gain.
- Before any /skill addition: gate the new skill on whether it survives an audit against the existing catalog.
What it does NOT do
- Does not edit any skill, prompt, router rule, or hook. Output is recommendations only.
- Does not benchmark — separate concern (see
docs/SOTA_2026.md).
- Does not propose new scaffolding. Its job is subtraction, not addition.
Why this is load-bearing
The Anthropic harness-design guide explicitly recommends: "strip away pieces that are no longer load-bearing to performance" as models improve. A harness that only grows is a harness that gets in its own way. /harness-audit is the codification of that subtraction discipline.