Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Surfaces relevant instincts during work. Use when starting a task to check if any learned behaviors apply.
.claude/skills/humanplane-instinct-apply/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-19 | ✗→✓ | ▲ Improved | -44% | 0% |
| case-17 | ✗→✓ | ▲ Improved | -9% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -40% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -7% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 6% | 0% |
You have learned behaviors. Use them.
bash# Read all personal instincts for f in .claude/homunculus/instincts/personal/*.md; do [ -f "$f" ] && echo "=== $(basename "$f") ===" && cat "$f" && echo done 2>/dev/null # Also check inherited instincts for f in .claude/homunculus/instincts/inherited/*.md; do [ -f "$f" ] && echo "=== $(basename "$f") ===" && cat "$f" && echo done 2>/dev/null
yaml--- trigger: "when [condition]" confidence: 0.7 domain: "code-style" --- # Name ## Action What to do ## Evidence Why this exists
When an instinct fires but the action feels wrong for the situation:
Instincts can be wrong. They're learned from patterns, and patterns have exceptions.
Don't read all instincts for every action. Keep relevant ones in working memory.
Quick domain check:
code-style instinctstesting instinctsgit instinctsdebugging instinctsBe efficient. Instincts are meant to help, not slow down.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-14 | fail→fail | 6,988 | 5,695 | -19% | 1 | 1 | 0% | 1,121 | 724 | -35% | 0 | 0 | — |
case-15 | pass→pass | 10,467 | 6,070 | -42% | 1 | 1 | 0% | 1,651 | 1,556 | -6% | 0 | 0 | — |
case-16 | pass→pass | 8,012 | 3,180 | -60% | 1 | 1 | 0% | 1,236 | 1,005 | -19% | 0 | 0 | — |
case-19 | fail→pass | 11,072 | 2,695 | -76% | 1 | 1 | 0% | 1,548 | 867 | -44% | 0 | 0 | — |
case-20 | fail→fail | 13,055 | 1,516 | -88% | 1 | 1 | 0% | 2,116 | 683 | -68% | 0 | 0 | — |
case-17 | fail→pass | 7,117 | 2,809 | -61% | 1 | 1 | 0% | 1,076 | 976 | -9% | 0 | 0 | — |
case-18 | pass→pass | 12,833 | 6,659 | -48% | 1 | 1 | 0% | 2,054 | 1,580 | -23% | 0 | 0 | — |
case-10 | fail→pass | 8,690 | 6,612 | -24% | 1 | 1 | 0% | 1,425 | 854 | -40% | 0 | 0 | — |
case-11 | pass→pass | 2,731 | 3,304 | +21% | 1 | 1 | 0% | 477 | 930 | +95% | 0 | 0 | — |
case-12 | fail→pass | 6,196 | 3,346 | -46% | 1 | 1 | 0% | 1,015 | 947 | -7% | 0 | 0 | — |
case-13 | fail→fail | 14,898 | 1,793 | -88% | 1 | 1 | 0% | 2,685 | 752 | -72% | 0 | 0 | — |
case-02 | fail→fail | 18,826 | 5,380 | -71% | 1 | 1 | 0% | 3,649 | 795 | -78% | 0 | 0 | — |
case-03 | fail→fail | 7,092 | 5,480 | -23% | 1 | 1 | 0% | 1,081 | 813 | -25% | 0 | 0 | — |
case-04 | pass→pass | 8,019 | 4,537 | -43% | 1 | 1 | 0% | 1,282 | 1,245 | -3% | 0 | 0 | — |
case-01 | fail→fail | 11,045 | 4,898 | -56% | 1 | 1 | 0% | 1,667 | 757 | -55% | 0 | 0 | — |
case-05 | pass→pass | 6,019 | 3,464 | -42% | 1 | 1 | 0% | 875 | 1,100 | +26% | 0 | 0 | — |
case-06 | pass→pass | 8,460 | 4,087 | -52% | 1 | 1 | 0% | 1,230 | 1,142 | -7% | 0 | 0 | — |
case-07 | fail→pass | 6,558 | 4,075 | -38% | 1 | 1 | 0% | 1,037 | 1,104 | +6% | 0 | 0 | — |
case-08 | pass→pass | 8,421 | 5,449 | -35% | 1 | 1 | 0% | 1,374 | 1,412 | +3% | 0 | 0 | — |
case-09 | fail→fail | 7,359 | 2,490 | -66% | 1 | 1 | 0% | 1,117 | 837 | -25% | 0 | 0 | — |
case-21 | pass→fail | 18,764 | 5,553 | -70% | 1 | 1 | 0% | 3,541 | 789 | -78% | 0 | 0 | — |
case-22 | pass→pass | 3,975 | 8,579 | +116% | 1 | 1 | 0% | 678 | 1,359 | +100% | 0 | 0 | — |
case-23 | pass→fail | 9,376 | 5,324 | -43% | 1 | 1 | 0% | 1,581 | 717 | -55% | 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 +13 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.