Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Core principle that the main agent is a coordinator, not an implementer. All work must be delegated to subagents.
.claude/skills/tzachbon-delegation-principle/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-03 | ✗→✓ | ▲ Improved | 35% | 0% |
| case-11 | ✗→✓ | ▲ Improved | 7% | 0% |
| case-20 | ✓→✓ | ＝ Same ✓ | -26% | 0% |
| case-21 | ✓→✓ | ＝ Same ✓ | -11% | 0% |
| case-22 | ✓→✓ | ＝ Same ✓ | -11% | 0% |
YOU MUST NEVER IMPLEMENT ANYTHING YOURSELF
The main agent (you) is a coordinator, not an implementer.
Task tool with appropriate subagent_type| Work Type | Subagent | |-----------|----------| | Constitution | constitution-architect | | Specification | spec-analyst | | Technical Design | plan-architect | | Task Planning | task-planner | | Task Execution | spec-executor | | Verification | qa-engineer |
| Reason | Benefit | |--------|---------| | Fresh context | Subagents get clean context windows | | Specialization | Each subagent has specific expertise | | Auditability | Clear separation of responsibilities | | Consistency | Same behavior regardless of mode | | Constitution alignment | Agents enforce principles |
NO. Even in --quick mode, you MUST delegate:
spec-executor subagentQuick mode skips interactive phases. Does NOT change delegation requirement.
textUser runs command ↓ Coordinator parses args ↓ Coordinator reads state ↓ Coordinator delegates via Task tool ↓ Subagent does ALL work ↓ Subagent returns result ↓ Coordinator reports to user ↓ Coordinator STOPS (unless quick mode)
After each phase completes:
awaitingApproval: true in stateException: --quick mode runs all phases without stopping.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 27,231 | 9,367 | -66% | 1 | 1 | 0% | 5,499 | 1,246 | -77% | 0 | 0 | — |
case-02 | fail→fail | 4,042 | 11,418 | +182% | 1 | 1 | 0% | 568 | 1,216 | +114% | 0 | 0 | — |
case-03 | fail→pass | 15,896 | 20,041 | +26% | 1 | 1 | 0% | 1,876 | 2,525 | +35% | 0 | 0 | — |
case-04 | fail→fail | 14,458 | 10,452 | -28% | 1 | 1 | 0% | 2,167 | 1,190 | -45% | 0 | 0 | — |
case-05 | fail→fail | 14,992 | 6,416 | -57% | 1 | 1 | 0% | 2,402 | 857 | -64% | 0 | 0 | — |
case-06 | fail→fail | 19,188 | 4,211 | -78% | 1 | 1 | 0% | 3,419 | 1,058 | -69% | 0 | 0 | — |
case-16 | fail→fail | 6,756 | 12,828 | +90% | 1 | 1 | 0% | 306 | 1,402 | +358% | 0 | 0 | — |
case-07 | fail→fail | 2,507 | 9,045 | +261% | 1 | 1 | 0% | 306 | 916 | +199% | 0 | 0 | — |
case-08 | fail→fail | 26,858 | 7,719 | -71% | 1 | 1 | 0% | 4,259 | 948 | -78% | 0 | 0 | — |
case-09 | fail→fail | 12,351 | 2,114 | -83% | 1 | 1 | 0% | 1,829 | 822 | -55% | 0 | 0 | — |
case-10 | fail→fail | 5,372 | 4,383 | -18% | 1 | 1 | 0% | 251 | 1,204 | +380% | 0 | 0 | — |
case-17 | fail→fail | 14,937 | 12,485 | -16% | 1 | 1 | 0% | 2,645 | 1,475 | -44% | 0 | 0 | — |
case-11 | fail→pass | 6,662 | 3,691 | -45% | 1 | 1 | 0% | 1,003 | 1,069 | +7% | 0 | 0 | — |
case-12 | fail→fail | 21,276 | 3,896 | -82% | 1 | 1 | 0% | 3,172 | 925 | -71% | 0 | 0 | — |
case-13 | fail→fail | 6,344 | 15,413 | +143% | 1 | 1 | 0% | 892 | 1,337 | +50% | 0 | 0 | — |
case-14 | fail→fail | 2,501 | 2,543 | +2% | 1 | 1 | 0% | 296 | 851 | +188% | 0 | 0 | — |
case-15 | fail→fail | 5,543 | 8,424 | +52% | 1 | 1 | 0% | 774 | 944 | +22% | 0 | 0 | — |
case-18 | fail→fail | 5,600 | 9,171 | +64% | 1 | 1 | 0% | 742 | 961 | +30% | 0 | 0 | — |
case-19 | fail→fail | 13,160 | 5,778 | -56% | 1 | 1 | 0% | 2,319 | 825 | -64% | 0 | 0 | — |
case-20 | pass→pass | 15,492 | 8,771 | -43% | 1 | 1 | 0% | 2,363 | 1,752 | -26% | 0 | 0 | — |
case-21 | pass→pass | 11,484 | 6,582 | -43% | 1 | 1 | 0% | 1,746 | 1,555 | -11% | 0 | 0 | — |
case-22 | pass→pass | 8,323 | 4,435 | -47% | 1 | 1 | 0% | 1,441 | 1,278 | -11% | 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 11 counted toward the lift figure. The other 11 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 +9 percentage points is the difference between those two pass rates over the 11 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.