← ClaudeAtlas

fanning-out-a-verified-read-only-swarmlisted

Read-only verified fan-out: split a question into atomic units, run them on lean agents (cheapest tier first, a mid tier when a unit can't split, the top tier only for judgment), verify every claim with a skeptic, double-verify with --compare, measure drift between runs. Scales to a whole corpus (one unit file per document). Invoking this skill (formerly /swarm) IS the Workflow opt-in. Triggers on: /swarm, swarm this, verify these claims, double verify, --compare, fact-check the report, audit read-only, drift between runs, reliability score.
sblattj/agent-chandlery · ★ 0 · Data & Documents · score 72
Install: claude install-skill sblattj/agent-chandlery
## More triggers swarm (old name) # fanning-out-a-verified-read-only-swarm **No Workflow tool in your client (OpenCode and others):** don't use this file; read `opencode.md` in this folder and follow it. Read-only fan-out that returns verified claims. Its sibling `fanning-out-parallel-writers-with-verified-changes` (formerly `/xx`) is the write-capable version, built on the same engine. Usage: `/fanning-out-a-verified-read-only-swarm <goal> [--compare] [--claims <file>] [--drift <runA> <runB>...] [--runs N]` Helpers live in `${SKILL_DIR:-~/.claude/skills/fanning-out-a-verified-read-only-swarm}`: `swarm.workflow.js` (the engine) and `swarm.py` (save, drift, ls). ## Do this 1. **Scout inline first** (a few greps or reads) to turn the goal into a work list. Don't fan out blind. 2. **Decompose into units**, making each one as small as possible: - `tier: haiku` is the default. One file, one ticket or one question per unit, answerable with shell reads. - `tier: sonnet` for a unit that can't be split without losing context (cross-file reasoning, a long diff). - `tier: opus` only for judgment that shouldn't be split (design calls, ambiguous intent). It's not banned, just last. - `needs_mcp: true` only when the unit needs MCP (a tracker MCP, a browser, a transcript server). It runs as `general-purpose`; everything else runs on `lean` (a shell + file tools agent type, about 10k vs 43k start tokens). If your client has no `lean` agent, define one or change `agentType