Freeze: preserve the exact baseline, promotion bank, graders, settings, thresholds, staged evidence budget, and stop rules. Include real failures.
Measure: inventory description, always-loaded body, each routed reference, full surface, and expected loaded tokens. Run scripts/measure_skill.py with --require-tokenizer or use an available tokenizer; record encoding and version. Label an unverified target-model mapping. Use provider receipts for execution input, cached input, output and reasoning; keep subsets separate. Character or word counts cannot qualify a token-saving candidate. Provide observed route frequencies.
Map: create a behavioral ledger: stable rule ID, decision changed, activation condition, owning file, dependent rules, positive case, and failure if lost. Separate catalog routing from post-load execution behavior.
Hypothesize: consider unchanged control, removal, and the shortest replacement per seam: concise semantic wording, established pattern names with necessary qualifiers, and telegraphic notation. Measure each against the unchanged text; symbols and abbreviations are not inherently fewer tokens. Prefer one independent variable. Reject noncompetitive variants by inspection; generating hypotheses does not require executing them.
Screen: run deterministic checks, then one observation per live variant on the most discriminating known failure. Reuse a condition-identical control observation. One hard regression may reject; one clean screen cannot promote.
Falsify: before replication, challenge survivors on a different failure mechanism, archetype, or grader. Run another trial only when its result can change the decision. Localize failures by restoring or splitting the changed seam.
Transfer: replicate survivors on the frozen required cells, including fresh held-out and compounded cases. Promote only when every cell passes and no material NFR regresses. Otherwise keep the baseline and report the smallest failing seam.
Apply: update the source, validate it, and prove it is byte-identical to the accepted candidate. Rerun the bank only if application transformed the bytes; report baseline, candidate, and installed hashes plus measured deltas.

Read evaluation protocol before designing the test bank. Read transformation mechanics when classifying or rewriting seams.

Optimization rules

Optimize catalog + always-loaded + routed-on-demand expected cost. Moving text to a reference is not compression when that reference always loads.
Preserve behavioral atoms, not necessarily their original sentences. A redundant, obsolete, default, contradictory, or harmful atom may be removed only by ablation.
Test both activation and restraint: cases where guidance is needed, irrelevant, misleading, and accumulated across multiple turns or deliverables.
Do not tune to evaluator nouns, fixtures, exact phrases, or one domain. Each retained mechanism must generalize to multiple archetypes or an explicit invariant.
Do not accept aggregate wins that hide a failed task cell or catastrophe.
Model-authored time estimates are not NFR evidence; use executor timestamps.
Keep raw prompts, outputs, traces, grades, invalid runs, and protocol deviations.
Stop when remaining candidates either fail quality gates or save less than the preregistered minimum practical token delta.

skill-compressor

Skill Compressor

Compression loop