Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Initialize a multi-agent swarm with anti-drift configuration. Use when starting a complex multi-file task that needs 3+ coordinated agents (feature implementation, refactor across modules, security audit). Skip for single-file edits or quick questions.
.claude/skills/ruvnet-swarm-init/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | -58% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -49% | 0% |
| case-05 | ✗→✓ | ▲ Improved | -70% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -68% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -51% | 0% |
Initialize a hierarchical swarm for coordinated multi-agent work.
Via MCP: mcp__plugin_ruflo-core_ruflo__swarm_init({ topology: "hierarchical", maxAgents: 8, strategy: "specialized" })
Or via CLI:
bashnpx @claude-flow/cli@latest swarm init --topology hierarchical --max-agents 8 --strategy specialized
Then spawn named agents in ONE message via Claude Code's Task tool with name: (for SendMessage addressability) and run_in_background: true (for parallel execution). Use EnterWorktree per agent for git-safe parallel work, and SendMessage for inter-agent coordination.
For larger teams (10+), use hierarchical-mesh topology:
bashnpx @claude-flow/cli@latest swarm init --topology hierarchical-mesh --max-agents 15 --strategy specialized
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 13,258 | 6,802 | -49% | 1 | 1 | 0% | 2,954 | 767 | -74% | 0 | 0 | — |
case-02 | fail→fail | 14,523 | 6,707 | -54% | 1 | 1 | 0% | 3,405 | 639 | -81% | 0 | 0 | — |
case-03 | fail→fail | 13,862 | 3,128 | -77% | 1 | 1 | 0% | 3,226 | 715 | -78% | 0 | 0 | — |
case-04 | fail→pass | 16,022 | 3,058 | -81% | 1 | 1 | 0% | 1,297 | 551 | -58% | 0 | 0 | — |
case-09 | fail→pass | 4,628 | 1,306 | -72% | 1 | 1 | 0% | 846 | 435 | -49% | 0 | 0 | — |
case-05 | fail→pass | 11,615 | 2,727 | -77% | 1 | 1 | 0% | 2,713 | 814 | -70% | 0 | 0 | — |
case-06 | pass→pass | 6,869 | 1,991 | -71% | 1 | 1 | 0% | 1,263 | 546 | -57% | 0 | 0 | — |
case-07 | pass→pass | 5,772 | 3,535 | -39% | 1 | 1 | 0% | 1,076 | 519 | -52% | 0 | 0 | — |
case-08 | pass→pass | 4,658 | 1,963 | -58% | 1 | 1 | 0% | 969 | 582 | -40% | 0 | 0 | — |
case-10 | pass→pass | 5,428 | 1,424 | -74% | 1 | 1 | 0% | 1,094 | 447 | -59% | 0 | 0 | — |
case-11 | fail→pass | 21,327 | 2,072 | -90% | 1 | 1 | 0% | 1,935 | 626 | -68% | 0 | 0 | — |
case-12 | fail→pass | 17,949 | 1,774 | -90% | 1 | 1 | 0% | 1,070 | 526 | -51% | 0 | 0 | — |
case-13 | fail→pass | 14,270 | 1,548 | -89% | 1 | 1 | 0% | 2,726 | 496 | -82% | 0 | 0 | — |
case-14 | fail→pass | 8,406 | 1,770 | -79% | 1 | 1 | 0% | 1,579 | 527 | -67% | 0 | 0 | — |
case-15 | fail→pass | 12,731 | 4,097 | -68% | 1 | 1 | 0% | 2,014 | 891 | -56% | 0 | 0 | — |
case-16 | fail→pass | 13,308 | 5,864 | -56% | 1 | 1 | 0% | 2,366 | 1,298 | -45% | 0 | 0 | — |
case-17 | fail→pass | 6,342 | 1,319 | -79% | 1 | 1 | 0% | 1,186 | 469 | -60% | 0 | 0 | — |
case-18 | fail→pass | 5,746 | 2,043 | -64% | 1 | 1 | 0% | 1,211 | 627 | -48% | 0 | 0 | — |
case-19 | fail→pass | 10,533 | 2,992 | -72% | 1 | 1 | 0% | 1,749 | 760 | -57% | 0 | 0 | — |
case-20 | pass→pass | 13,763 | 18,827 | +37% | 1 | 1 | 0% | 2,584 | 3,416 | +32% | 0 | 0 | — |
case-21 | pass→pass | 8,619 | 4,905 | -43% | 1 | 1 | 0% | 1,791 | 1,042 | -42% | 0 | 0 | — |
case-22 | pass→pass | 11,672 | 10,211 | -13% | 1 | 1 | 0% | 2,350 | 1,949 | -17% | 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 19 counted toward the lift figure. The other 3 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 +55 percentage points is the difference between those two pass rates over the 19 comparable cases. 1 case got worse with the skill loaded, and it is 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.