| name | nw-ab-optimize-skill |
| description | PROCEDURE — optimize one requested existing skill. Measures multi-trigger, duplication, and resident-context debt before editing; preserves its name as the lean core and behavioral projections. |
| user-invocable | false |
nw-ab-optimize-skill (PROCEDURE)
Kind: PROCEDURE | One job: optimize one existing skill | One trigger: an existing skill is requested for optimization.
Eligibility
- Measure debt — Record source lines, words, bytes; classify every rule and firing condition; identify duplication and resident-context cost. Compute
static_resident; compute expected_runtime only from evidenced trigger probabilities, otherwise INDETERMINATE; compute worst_runtime. Use “tokens” only for exact compatible token measurements, otherwise report INDETERMINATE. Stop: a small coherent single-trigger skill; return NOT_ELIGIBLE without edits.
- Protect behavior — Inventory the skill's public name, global invariants, referenced callers, textual pins, public/private fan-out, and existing behavioral projections. Stop: baseline red, a pin cannot be preserved, or a public fan-out is missing.
Deterministic workflow
At execution start create these as TaskCreate items and run in order.
- Map ownership — Assign every original rule exactly one owner: original lean core, existing reused skill, or one extracted skill. Reuse first. Stop: duplicate or unowned rule.
- Partition triggers — State a concrete trigger for every extracted procedure or knowledge skill. The triggers together cover the original trigger-space, without overlap. Keep one-job/one-trigger skills intact. Stop: trigger gap or overlap.
- Construct first — Apply this order: remove, relax, route, construct, then consider an existing check. When an LLM still authors formal bytes, templates, grammars, reviewers, and validators are checks, not construction; prefer a deterministic sole-writer constructor or compiler that accepts semantic input once. Do not add enforcement machinery (validator, hook, gate, test, or mandatory artifact) through this procedure; a deterministic constructor/compiler that replaces LLM formal-byte authoring is allowed.
- Recompose — Preserve the original skill name as the lean core. Keep its routing, global invariants, and public contract there; extract only distinct-trigger procedures or knowledge. Wire every extracted skill from the core; no orphan or name change.
- Measure outcome — Re-measure core lines, words, bytes,
static_resident, expected_runtime when evidenced, and worst_runtime. Retain a step only with evidence that it improves speed, resident/expected context cost, or outcome quality without silently worsening another; otherwise remove or skip it. Claim improvement only for a measured reduction; otherwise label the result STRUCTURAL_ONLY and name no efficiency gain. Stop: resident cost worsens without an explicit decision.
- Verify once — Batch edits, run the existing relevant validation once, and execute the existing behavioral projections once. Do not create new validation. Stop: behavior changes or an existing projection is red.
- Handoff — Return compact evidence only: ownership map; trigger partition; before/after metrics; executed projections;
independent-review-needed; residual INDETERMINATE. Do not create a mandatory report artifact.
Invariants
- Original skill name, public contract, and behavioral projections are preserved.
- A small coherent skill is
NOT_ELIGIBLE; line count alone never justifies extraction.
- Resident metrics name their scope. Static resident is always measured; expected runtime is
INDETERMINATE without probability evidence.
- This procedure creates no ceremony. Construction changes the producer; a later formal-byte authoring step is not construction.
- Prefer deletion over new machinery. Structural-only benefit is never sold as token saving.
Success criteria
