meta-skills

Audit and score a repo's agent context file (AGENTS.md, CLAUDE.md, .hermes.md, .cursorrules) for how Hermes actually loads it: the 20KB truncation cap, one-file-wins priority, injection-scanner blocks, SOUL-vs-AGENTS placement, plus a Karpathy 12-rule completeness rubric and per-line instruction lints. Use when the user says audit my AGENTS.md, review this CLAUDE.md, check my context file, is my AGENTS.md too big, why isn't my AGENTS.md being followed, grade my repo instructions, or points at an AGENTS.md/CLAUDE.md and asks if it's good. For auditing a SKILL.md instead, use skill-audit.

Audit and score a SOUL.md (Hermes agent identity file) or any agent persona / system-prompt / operator-instruction doc, against how Hermes loads it plus attention research. Use when the user shares a SOUL.md, persona file, agent constitution, or system-prompt and asks to review/grade/audit it, 'is this good', or 'compare to best practice'. Not for skills (use skill-audit).

How to clean up duplicate/untracked skills in the skills repo

Eval-driven skill optimizer: runs a skill repeatedly, scores outputs against binary evals, mutates the prompt via structured edits, and keeps only changes that improve a held-out validation score (cross-model optimizer/target, three-way splits, golden cases, checkpoint resume). Use when: optimize/improve this skill, make this skill better, run autoresearch on, self-improve skill, benchmark/eval my skill, run evals on.