Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Compare a claimed state with observable
.claude/skills/boshu2-reality-check/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-21 | ✗→✓ | ▲ Improved | 109% | 0% |
| case-05 | ✓→✗ | ▼ Worse | -50% | 0% |
| case-08 | ✓→✗ | ▼ Worse | -44% | 0% |
| case-11 | ✓→✗ | ▼ Worse | 5% | 0% |
| case-12 | ✓→✗ | ▼ Worse | 319% | 0% |
Compare an explicit claim with observable evidence. Cite every confirmed or missing behavior with a file, command result, or artifact. Separate:
When the claim is a completion or status claim, audit it against the stated goals, not against what happens to exist. Enumerate every goal in the vision, plan, or intent source and give each a disposition: confirmed with evidence, concrete gap, or unverifiable. The audit is complete only when every stated goal carries a disposition; full coverage of the built surface alone proves nothing about completion. The named failure mode is built-world bias: auditing only the code that exists, so goals nobody started never surface as gaps.
When the same check runs across multiple passes or sessions, freeze the exact question wording before the first pass and ask it identically in every pass; record the frozen wording in the report. A pass that answers a reworded question starts a new baseline — comparing it against earlier passes is the drifting-rubric failure mode, and its answer does not count as a repeated measurement.
When invoked during planning, compare the currently planned scope against the originally stated goal. Planned work that cannot be traced to a stated goal is reported as an escalation gap, exactly like a missing behavior. Reality Check reports the escalation; the caller decides whether the ambition or the stated goal changes.
.agents/scratch/reality-check/<run-id>/.reality-check-report.json.reality-check-report.v1 JSON — the checked claim, one finding perconfirmed behavior, concrete gap, incomplete-evidence item, or changed assumption (each with cited evidence), and, for a completion or status claim, the goal-by-goal coverage disposition. It carries no verdict, readiness, or PASS field; the validator rejects one.
skills/reality-check/scripts/validate-output.sh <reality-check-report.json>.
If the claim cannot be tested against any observable evidence, report it as incomplete-evidence with the missing artifact named — never resolve an untestable claim as confirmed.
Return the report to the caller. Plan may use concrete gaps to refine the existing bead or caller intent. Reality Check reports observations; it does not mint a verdict or PASS of any version, create work, schedule, claim, implement, validate, retry, or deliver.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 13,612 | 8,842 | -35% | 1 | 1 | 0% | 287 | 1,047 | +265% | 0 | 0 | — |
case-02 | fail→fail | 4,671 | 6,860 | +47% | 1 | 1 | 0% | 244 | 784 | +221% | 0 | 0 | — |
case-03 | fail→fail | 4,534 | 5,475 | +21% | 1 | 1 | 0% | 230 | 874 | +280% | 0 | 0 | — |
case-04 | fail→fail | 7,731 | 7,278 | -6% | 1 | 1 | 0% | 1,321 | 938 | -29% | 0 | 0 | — |
case-05 | pass→fail | 10,463 | 6,926 | -34% | 1 | 1 | 0% | 1,905 | 947 | -50% | 0 | 0 | — |
case-06 | fail→fail | 13,771 | 10,220 | -26% | 1 | 1 | 0% | 2,181 | 1,059 | -51% | 0 | 0 | — |
case-07 | fail→fail | 14,299 | 7,443 | -48% | 1 | 1 | 0% | 2,059 | 968 | -53% | 0 | 0 | — |
case-08 | pass→fail | 10,110 | 7,671 | -24% | 1 | 1 | 0% | 1,823 | 1,017 | -44% | 0 | 0 | — |
case-09 | fail→fail | 8,232 | 25,967 | +215% | 1 | 1 | 0% | 220 | 4,580 | +1982% | 0 | 0 | — |
case-10 | fail→fail | 2,371 | 16,965 | +616% | 1 | 1 | 0% | 308 | 3,601 | +1069% | 0 | 0 | — |
case-11 | pass→fail | 12,297 | 10,475 | -15% | 1 | 1 | 0% | 2,098 | 2,197 | +5% | 0 | 0 | — |
case-12 | pass→fail | 6,033 | 18,103 | +200% | 1 | 1 | 0% | 963 | 4,036 | +319% | 0 | 0 | — |
case-13 | fail→fail | 12,733 | 4,857 | -62% | 1 | 1 | 0% | 2,115 | 893 | -58% | 0 | 0 | — |
case-14 | fail→fail | 4,413 | 5,717 | +30% | 1 | 1 | 0% | 209 | 892 | +327% | 0 | 0 | — |
case-15 | fail→fail | 3,807 | 4,735 | +24% | 1 | 1 | 0% | 164 | 784 | +378% | 0 | 0 | — |
case-16 | fail→fail | 19,055 | 7,739 | -59% | 1 | 1 | 0% | 3,818 | 1,158 | -70% | 0 | 0 | — |
case-17 | pass→fail | 17,044 | 14,185 | -17% | 1 | 1 | 0% | 1,296 | 924 | -29% | 0 | 0 | — |
case-18 | fail→fail | 6,208 | 6,271 | +1% | 1 | 1 | 0% | 211 | 989 | +369% | 0 | 0 | — |
case-19 | fail→fail | 11,415 | 6,358 | -44% | 1 | 1 | 0% | 1,370 | 977 | -29% | 0 | 0 | — |
case-20 | pass→pass | 5,511 | 16,194 | +194% | 1 | 1 | 0% | 825 | 2,594 | +214% | 0 | 0 | — |
case-21 | fail→pass | 9,061 | 12,159 | +34% | 1 | 1 | 0% | 1,400 | 2,928 | +109% | 0 | 0 | — |
case-22 | pass→fail | 11,142 | 6,936 | -38% | 1 | 1 | 0% | 1,934 | 999 | -48% | 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, and 5 counted toward the lift figure. The other 17 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 -50 percentage points is the difference between those two pass rates over the 5 comparable cases. 9 cases got worse with the skill loaded, and they are included in that figure.
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.