Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Distill repeated, evidence-backed expertise
.claude/skills/boshu2-operationalize/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 26% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 51% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -10% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 56% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -26% | 0% |
Turn repeated, cited expertise into a proposal for a reusable artifact.
authoritative source, subject to the three-instance floor below when the proposal abstracts a rule.
examples, and evidence.
certificate, ledger, dashboard, matrix, meta-report, readiness review, speculative check, skill, or workflow must name its concrete consumer, the subject or release decision it gates, the observed defect class justifying it, and its deletion condition. Code or process introduced solely to consume the artifact does not qualify. If any answer is missing, propose no artifact and redirect to the caller-requested subject. Minimal integrity or recovery state is allowed only when necessary to prevent a named evidence-loss or corruption mode.
caller-owned workflow.
or deletion condition.
the caller asks for a durable artifact, write it under .agents/scratch/operationalize/ first and return the path; the proposal is advisory either way.
A rule needs three real occurrences before it may be abstracted. Count only occurrences that actually happened and can be cited — sessions, diffs, verdicts, or artifacts that resolve in this repository — not hypothetical cases or restatements of one event. With one or two occurrences, propose a quote-anchored reference note instead and stop short of a rule. An explicit authoritative source may substitute for occurrences only when the proposal transcribes that source rather than generalizing beyond it. The named failure mode is premature abstraction: a rule minted from a single vivid incident that encodes the incident's accidents as policy.
Every proposed rule carries a reapply proof: a demonstration that the rule, as written, reproduces the correct decision on at least one of its source occurrences without extra context. If applying the drafted rule to its own source moment requires unwritten judgment, the rule is not yet operational — tighten the wording until the reapply succeeds, or downgrade the proposal to a reference. When the proposal creates process, the reapply proof must also show that the creation gate returns the correct create-or-drop decision. No reapply proof, no rule.
Tie each rule to its source moments with a quote bank: for every counted occurrence, a short verbatim quote or command/output excerpt plus a locally resolving citation (repo path, .agents/ao digest, or session artifact). An occurrence that cannot be quoted and cited does not count toward the three-instance floor. Anchors let a later reader test whether the rule still matches what actually happened, instead of trusting the abstraction.
Operationalize does not create tracker work, promote policy, start a factory, validate its own output, or control another invocation. The proposal is advisory: adopting it into a skill, deterministic check, reference, or workflow is a separate, caller-selected step — skill-builder, workflow-builder, or a fresh RPI — never performed here. The proposal cannot promote itself, and process-only output earns no capability credit.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→pass | 16,419 | 14,857 | -10% | 1 | 1 | 0% | 2,737 | 3,442 | +26% | 0 | 0 | — |
case-02 | fail→fail | 16,714 | 11,616 | -31% | 1 | 1 | 0% | 2,691 | 2,863 | +6% | 0 | 0 | — |
case-03 | fail→pass | 19,654 | 22,656 | +15% | 1 | 1 | 0% | 3,055 | 4,621 | +51% | 0 | 0 | — |
case-04 | fail→pass | 15,676 | 8,050 | -49% | 1 | 1 | 0% | 2,285 | 2,047 | -10% | 0 | 0 | — |
case-05 | fail→pass | 13,254 | 14,381 | +9% | 1 | 1 | 0% | 2,053 | 3,203 | +56% | 0 | 0 | — |
case-06 | pass→pass | 16,353 | 20,313 | +24% | 1 | 1 | 0% | 2,743 | 4,157 | +52% | 0 | 0 | — |
case-07 | fail→pass | 18,820 | 7,106 | -62% | 1 | 1 | 0% | 2,819 | 2,081 | -26% | 0 | 0 | — |
case-08 | fail→pass | 13,579 | 8,306 | -39% | 1 | 1 | 0% | 2,148 | 2,112 | -2% | 0 | 0 | — |
case-17 | pass→pass | 13,337 | 10,733 | -20% | 1 | 1 | 0% | 2,307 | 2,494 | +8% | 0 | 0 | — |
case-09 | pass→pass | 14,301 | 7,030 | -51% | 1 | 1 | 0% | 2,285 | 1,935 | -15% | 0 | 0 | — |
case-10 | pass→pass | 12,432 | 6,762 | -46% | 1 | 1 | 0% | 1,800 | 1,880 | +4% | 0 | 0 | — |
case-11 | fail→fail | 21,790 | 22,658 | +4% | 1 | 1 | 0% | 3,648 | 4,887 | +34% | 0 | 0 | — |
case-12 | pass→pass | 10,338 | 12,432 | +20% | 1 | 1 | 0% | 1,931 | 3,014 | +56% | 0 | 0 | — |
case-13 | pass→pass | 12,500 | 13,417 | +7% | 1 | 1 | 0% | 1,878 | 3,104 | +65% | 0 | 0 | — |
case-14 | fail→pass | 15,353 | 4,485 | -71% | 1 | 1 | 0% | 2,370 | 1,495 | -37% | 0 | 0 | — |
case-15 | pass→pass | 10,425 | 9,566 | -8% | 1 | 1 | 0% | 1,740 | 2,315 | +33% | 0 | 0 | — |
case-16 | pass→pass | 15,273 | 14,228 | -7% | 1 | 1 | 0% | 2,416 | 3,299 | +37% | 0 | 0 | — |
case-18 | pass→pass | 12,447 | 3,759 | -70% | 1 | 1 | 0% | 1,912 | 1,466 | -23% | 0 | 0 | — |
case-19 | fail→pass | 12,009 | 7,111 | -41% | 1 | 1 | 0% | 1,904 | 1,967 | +3% | 0 | 0 | — |
case-20 | fail→pass | 16,972 | 7,287 | -57% | 1 | 1 | 0% | 1,980 | 2,016 | +2% | 0 | 0 | — |
case-21 | pass→pass | 8,542 | 7,485 | -12% | 1 | 1 | 0% | 1,327 | 2,069 | +56% | 0 | 0 | — |
case-22 | fail→pass | 27,096 | 9,280 | -66% | 1 | 1 | 0% | 5,747 | 2,343 | -59% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +45 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.