Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Mandatory protocol that prevents AI agents from fabricating information. Forces best practice search before every action — coding, design, content, pricing, naming, architecture, everything. Every decision must have a citation (source name + URL + direct quote). Based on Anthropic official hallucination reduction techniques.
.claude/skills/majiayu000-investigate-before-acting/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 95% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 163% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 60% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 23% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 2219% | 0% |
Check if SOUL.md contains the marker <!-- investigate-before-acting: installed -->.
{baseDir}/references/reasoning.md — understand WHY each rule exists{baseDir}/references/soul-patch.md — append its contents to SOUL.md (create SOUL.md if it does not exist){baseDir}/references/agents-patch.md — append its contents to AGENTS.md (skip if AGENTS.md does not exist).claude/CLAUDE.md exists in the current project, read {baseDir}/references/claude-patch.md and append its contentsソース: [Title] (URL) / 核心の引用: 「[direct quote]」| Rule | Why | |------|-----| | No questions to the user | The user doesn't know the best practice. Search engines do. Ask them instead. | | No options / no "A or B?" | Sufficient research converges to one answer. Two options = insufficient research. | | No originals | The success formula exists. Remove yourself from the equation. Copy. 100%. | | "No BP exists" is impossible | Buddhism answers "how to end suffering." Everything more concrete has an answer too. | | Generalize every lesson | Narrow lessons prevent one failure. Generalized principles prevent all similar failures. |
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 13,605 | 15,614 | +15% | 1 | 1 | 0% | 305 | 917 | +201% | 0 | 0 | — |
case-02 | fail→fail | 31,314 | 29,218 | -7% | 1 | 1 | 0% | 4,492 | 1,045 | -77% | 0 | 0 | — |
case-03 | fail→fail | 53,156 | 55,644 | +5% | 1 | 1 | 0% | 8,240 | 8,866 | +8% | 0 | 0 | — |
case-04 | fail→pass | 18,580 | 27,514 | +48% | 1 | 1 | 0% | 2,422 | 4,714 | +95% | 0 | 0 | — |
case-05 | fail→pass | 18,963 | 32,612 | +72% | 1 | 1 | 0% | 2,455 | 6,467 | +163% | 0 | 0 | — |
case-06 | fail→fail | 17,589 | 6,909 | -61% | 1 | 1 | 0% | 2,629 | 1,089 | -59% | 0 | 0 | — |
case-07 | fail→pass | 14,289 | 18,684 | +31% | 1 | 1 | 0% | 2,392 | 3,816 | +60% | 0 | 0 | — |
case-08 | fail→pass | 12,411 | 8,948 | -28% | 1 | 1 | 0% | 1,606 | 1,969 | +23% | 0 | 0 | — |
case-09 | fail→fail | 4,362 | 5,394 | +24% | 1 | 1 | 0% | 183 | 1,055 | +477% | 0 | 0 | — |
case-10 | fail→pass | 4,776 | 26,057 | +446% | 1 | 1 | 0% | 211 | 4,893 | +2219% | 0 | 0 | — |
case-11 | fail→fail | 18,640 | 18,930 | +2% | 1 | 1 | 0% | 2,952 | 3,971 | +35% | 0 | 0 | — |
case-12 | pass→pass | 16,012 | 16,792 | +5% | 1 | 1 | 0% | 2,039 | 3,569 | +75% | 0 | 0 | — |
case-13 | fail→pass | 14,495 | 17,916 | +24% | 1 | 1 | 0% | 2,306 | 3,942 | +71% | 0 | 0 | — |
case-14 | fail→pass | 12,802 | 14,945 | +17% | 1 | 1 | 0% | 2,041 | 3,350 | +64% | 0 | 0 | — |
case-15 | fail→pass | 21,711 | 19,288 | -11% | 1 | 1 | 0% | 3,191 | 4,030 | +26% | 0 | 0 | — |
case-16 | fail→pass | 23,778 | 37,079 | +56% | 1 | 1 | 0% | 3,474 | 7,754 | +123% | 0 | 0 | — |
case-17 | fail→pass | 5,324 | 20,283 | +281% | 1 | 1 | 0% | 878 | 4,449 | +407% | 0 | 0 | — |
case-18 | pass→pass | 10,975 | 19,870 | +81% | 1 | 1 | 0% | 1,671 | 4,189 | +151% | 0 | 0 | — |
case-19 | pass→fail | 1,626 | 12,457 | +666% | 1 | 1 | 0% | 247 | 2,904 | +1076% | 0 | 0 | — |
case-20 | pass→pass | 2,489 | 4,865 | +95% | 1 | 1 | 0% | 403 | 1,459 | +262% | 0 | 0 | — |
case-21 | pass→pass | 4,200 | 13,926 | +232% | 1 | 1 | 0% | 881 | 3,300 | +275% | 0 | 0 | — |
case-22 | fail→pass | 13,209 | 18,287 | +38% | 1 | 1 | 0% | 1,676 | 3,970 | +137% | 0 | 0 | — |
case-23 | fail→pass | 17,316 | 17,246 | -0% | 1 | 1 | 0% | 2,842 | 3,950 | +39% | 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 18 counted toward the lift figure. The other 5 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 +48 percentage points is the difference between those two pass rates over the 18 comparable cases. 2 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.