Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Run one bounded Plan, Implement, and fresh
.claude/skills/boshu2-rpi/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | 1654% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -39% | 0% |
| case-03 | ✗→✓ | ▲ Improved | 128% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 250% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 61% | 0% |
Run one experiment from the caller's existing intent source through three responsibilities and stop:
textanti-ceremony guard -> Plan -> Implement -> fresh Validate -> report
On CONTINUE, the core path remains Plan -> Implement -> fresh Validate -> report. RPI invokes the guard exactly once before Plan. It preserves the original intent and dispatches each core phase at most once. It does not own retries, budgets, queues, claims, leases, Git, delivery, release, closure, or the caller's next decision.
The pure scripts/run_once.py reference behavior makes the dispatch and stop semantics executable without Git, ao, or a tracker.
RPI activates for any request shaped as plan-execute-verify work — orchestration, worker delegation, "execute this plan", or an explicit Plan -> Implement -> Validate ask — whenever the goal includes changing the subject. The caller does not have to name RPI. Research-, audit-, and review-only delegation is not RPI admission: it produces evidence for a caller, has no implementation candidate, and never earns a verdict.
Once the caller has accepted a plan — including a duel or design synthesis — Plan is closed for that intent. Every subsequent lane must return implementation evidence: diffs, commits, test results, or factual receipts. Dispatching another planning, audit, or review lane over the same intent requires new explicit caller authorization; a review comment is never that authorization by itself.
outcome, proposed process work, remaining proof, and stop condition. On STOP, dispatch no core phase, report NOT_PLANNED with the guard's one-sentence reason, and stop. On CONTINUE, proceed without adding an artifact, retry, repair, delivery, tracker, or Git action.
source needs shaping; Plan updates the same source or proposes an amendment. It creates no AgentOps packet. Preserve a durable caller-owned source by reference and digest; only when no durable source exists does the runtime snapshot the exact resolved source bytes under their digest before dispatching Implement or a fresh Validate context. If usable intent cannot be established, report NOT_PLANNED and stop.
experiment; the runtime derives subject identity and check receipts. If no subject is built, report NOT_BUILT and stop.
the intent reference and digest, exact subject manifest, factual receipts, validator identity, and freshness attestation.
verdict.v2 only when the caller requests machine-readable evidence or a declared downstream consumer requires it. Stop regardless of PASS, FAIL, or NOT_PROVEN.
NOT_PLANNED and NOT_BUILT are report statuses, never semantic verdicts. A caller may revise the bead or caller intent and start a new invocation. RPI never creates a parallel revision artifact or selects the next work itself.
The hard anti-ceremony dependency owns the quick guard and its explicit-only full honesty audit. RPI does not duplicate that judgment or turn each component, gate failure, or specialist comment into a new planning artifact. A terminal caller goal may remain one bounded experiment across several source owners when they serve one outcome and one acceptance boundary.
If control artifacts or fresh-validation cycles are multiplying faster than implementation evidence, stop dispatching more lanes. Return to one outcome-level intent and continue with targeted deterministic checks, reserving the full integration check and fresh validation for the frozen subject. This changes orchestration cost, never acceptance, exact identity, fail-closed scope, or validation authority.
The spiral breaker fires when two consecutive control artifacts (plans, audits, reviews, prompts, reports) contain no new implementation evidence. Terminate the run and report NOT_BUILT when no implementation subject exists; when a subject exists, stop and report its current status without dispatching another lane or repair revision. RPI owns no lane budget, repair budget, or retry policy.
Delegate with minimal context: a lane receives the frozen intent reference and the established facts it needs, never the orchestrator's full conversation history. If a lane cannot proceed from the intent alone, report that the plan failed the fresh-context test and stop; do not pad it with chat transcript or start another planning lane without explicit caller authorization.
Lanes whose write scopes share a regen surface (the same generated outputs, mirrors, or manifests) serialize; only lanes with disjoint source scopes and disjoint regen surfaces may run in parallel.
STOP dispatches none of Plan,Implement, or Validate, while CONTINUE preserves their order.
NOT_PROVEN.
write_scope makes the verdict FAIL.explicit freshness attestation.
adapters are caller-selected. They do not alter phase order or core outcomes. When a factory adapter is selected, work enters it through that factory's coordinator (for Gas City, the Mayor — see using-gc); RPI hands over intent and never dispatches factory runs itself.
this invocation.
RPI has one required report surface and one optional representation:
language. This is the default assistant response.
rpi-report.v1 objectonly when the caller requests machine-readable evidence or a declared adapter consumes it. The schema ships in a repo checkout at schemas/rpi-report.v1.schema.json; the minimal required shape is:
json { "schema_version": "rpi-report.v1", "status": "PASS", "intent_ref": "<durable-source-ref-or-fallback-snapshot-ref>", "acceptance_digest": "<64-hex-char-sha256-or-null>", "subject_manifest_digest": "<64-hex-char-sha256-or-null>", "verdict_ref": "<verdict-location-or-null>", "verdict_digest": "<64-hex-char-sha256-or-null>", "checked": ["<criterion satisfied by evidence>"], "not_checked": ["<criterion not covered>"] }
intent_ref remains required: it names the durable caller-owned source when one exists, otherwise the content-addressed fallback snapshot. status is one of PASS | FAIL | NOT_PROVEN | NOT_PLANNED | NOT_BUILT; the three digest fields, when present, are 64-character lowercase hex SHA-256 strings; checked and not_checked are arrays of strings. All nine keys are required (use null for an inapplicable ref or digest), and no additional properties are allowed.
Lead the interactive response with the status and one sentence stating the caller-visible outcome. Lead with the subject, not the process: production paths changed, commits, test results, and acceptance criteria satisfied or remaining. A rising artifact count over an unchanged subject is a stop signal, not progress. Follow with only the strongest proof, any material unchecked scope, and a clickable verdict reference when one exists. Name why no subject exists for NOT_PLANNED or NOT_BUILT; for a guard STOP, use its one-sentence reason. Keep the response to one short paragraph or at most four bullets.
When no machine artifact was requested, do not create a hidden one. Raw digests, schema fields, and exhaustive check lists stay out of the interactive response unless an integrity failure makes one necessary to explain the result.
Do not append a next action. The caller owns continuation.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-11 | pass→pass | 7,376 | 2,181 | -70% | 1 | 1 | 0% | 1,260 | 2,271 | +80% | 0 | 0 | — |
case-01 | fail→pass | 4,281 | 12,852 | +200% | 1 | 1 | 0% | 219 | 3,841 | +1654% | 0 | 0 | — |
case-02 | fail→pass | 31,117 | 12,325 | -60% | 1 | 1 | 0% | 6,423 | 3,896 | -39% | 0 | 0 | — |
case-03 | fail→pass | 12,165 | 14,578 | +20% | 1 | 1 | 0% | 2,162 | 4,924 | +128% | 0 | 0 | — |
case-04 | fail→fail | 7,963 | 15,519 | +95% | 1 | 1 | 0% | 850 | 3,507 | +313% | 0 | 0 | — |
case-05 | pass→pass | 18,848 | 12,153 | -36% | 1 | 1 | 0% | 2,875 | 3,713 | +29% | 0 | 0 | — |
case-06 | pass→pass | 9,562 | 7,527 | -21% | 1 | 1 | 0% | 1,486 | 2,926 | +97% | 0 | 0 | — |
case-07 | fail→pass | 4,590 | 2,747 | -40% | 1 | 1 | 0% | 660 | 2,307 | +250% | 0 | 0 | — |
case-08 | fail→pass | 10,439 | 4,564 | -56% | 1 | 1 | 0% | 1,750 | 2,814 | +61% | 0 | 0 | — |
case-09 | pass→pass | 10,314 | 2,157 | -79% | 1 | 1 | 0% | 1,609 | 2,329 | +45% | 0 | 0 | — |
case-10 | pass→pass | 8,934 | 2,450 | -73% | 1 | 1 | 0% | 1,461 | 2,330 | +59% | 0 | 0 | — |
case-12 | pass→pass | 8,245 | 2,557 | -69% | 1 | 1 | 0% | 1,224 | 2,440 | +99% | 0 | 0 | — |
case-13 | pass→pass | 7,523 | 2,247 | -70% | 1 | 1 | 0% | 1,114 | 2,297 | +106% | 0 | 0 | — |
case-14 | pass→pass | 6,658 | 2,796 | -58% | 1 | 1 | 0% | 981 | 2,430 | +148% | 0 | 0 | — |
case-15 | pass→pass | 3,996 | 2,070 | -48% | 1 | 1 | 0% | 619 | 2,307 | +273% | 0 | 0 | — |
case-16 | fail→pass | 12,403 | 3,174 | -74% | 1 | 1 | 0% | 1,883 | 2,433 | +29% | 0 | 0 | — |
case-17 | fail→pass | 8,808 | 3,276 | -63% | 1 | 1 | 0% | 1,417 | 2,539 | +79% | 0 | 0 | — |
case-18 | fail→pass | 12,473 | 2,969 | -76% | 1 | 1 | 0% | 1,928 | 2,398 | +24% | 0 | 0 | — |
case-19 | fail→pass | 6,188 | 2,309 | -63% | 1 | 1 | 0% | 873 | 2,280 | +161% | 0 | 0 | — |
case-20 | fail→pass | 10,467 | 2,595 | -75% | 1 | 1 | 0% | 1,541 | 2,375 | +54% | 0 | 0 | — |
case-21 | fail→pass | 12,519 | 2,356 | -81% | 1 | 1 | 0% | 1,661 | 2,297 | +38% | 0 | 0 | — |
case-22 | fail→fail | 6,284 | 3,070 | -51% | 1 | 1 | 0% | 933 | 2,515 | +170% | 0 | 0 | — |
case-23 | fail→pass | 7,587 | 2,362 | -69% | 1 | 1 | 0% | 1,141 | 2,251 | +97% | 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. 23 cases were attempted, and 21 counted toward the lift figure. The other 2 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +52 percentage points is the difference between those two pass rates over the 21 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.