| name | detoks-skill |
| description | Use when optimizing Agent Skills for lower token usage, creating compact companion skills, splitting skill instructions into references, extracting deterministic workflow scripts, or adding eval-backed skill compression |
detoks-skill
Make token-optimized companion Agent Skills. Preserve original. Generate compact root, move detail to references/, extract deterministic work to scripts/, validate with AgentEvals.
Load Map
references/workflow.md: end-to-end source inspect, companion generation, token budget report.references/compression-strategies.md: TOC split, detok/caveman pass, preservation rules.references/script-extraction.md: deterministic workflow detection and TS script extraction.references/evals.md: AgentEvals EVAL.yaml contract and acceptance checks.
Scripts
scripts/analyze-skill.ts: token hotspots + deterministic candidates.scripts/render-optimized-skill.ts: write companion skill without mutating source.scripts/validate-optimized-skill.ts: frontmatter, references, token ratio, preservation checks.scripts/evolve-candidates.ts: optimize candidates through real Microsoft Trace (trace-opt) or Agent Lightning (agentlightning) backend.scripts/optimize-agent-frontmatter.ts: compact frontmatter context, preserve declarations.
Rules
- Never overwrite source skill.
- Keep optimized root as table of contents.
- Preserve code blocks, paths, commands, frontmatter, URLs, IDs exactly.
- Put source-heavy guidance in
references/; put deterministic repeated steps in scripts/.
- Prefer repo
detoks-prompt CLI for reference prose compression when available.
- Use
evals/EVAL.yaml as quality gate before reporting success.
