菜单
Initialize a batch processing job for one or more SWARM experiment runs. Sets up the queue, calculates claim numbering, creates the extract task. The entry point for pipeline processing of new runs. Triggers on "/seed", "/seed [run-folder]", "start processing", "queue this run".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Find connections between claims and update topic maps. Requires semantic judgment to identify genuine evidence relationships. Use after /extract creates claims, when exploring connections, or when a topic map needs synthesis. Triggers on "/cross-link", "/cross-link [claim]", "find connections", "update topic maps", "connect these claims".
Extract structured claims from SWARM experiment runs, sweep results, and red-team reports. Comprehensive extraction is the default — every governance finding, adversarial pattern, and parameter sensitivity result gets extracted. Zero extraction from a domain-relevant run is a BUG. Triggers on "/extract", "/extract [file]", "extract findings", "mine this run", "process this experiment".
Explore and analyze the SWARM knowledge graph — detect orphan claims, find dense connection clusters, trace evidence chains, identify synthesis opportunities. Triggers on "/graph", "graph analysis", "find orphans", "what connects to X".
Research a topic relevant to SWARM AI safety and grow the vault with findings. Queries external sources, synthesizes into claim candidates, follows the extraction pipeline. Triggers on "/learn", "/learn [topic]", "research this", "find papers on".
Surface the next highest-priority action from the SWARM synthesis pipeline queue and maintenance conditions. Recommends what to work on based on queue state, pending observations, stale claims, and research threads. Triggers on "/next", "what should I work on", "next action", "what's pending".
Run the full SWARM synthesis pipeline on one or more experiment runs — seed, extract, cross-link, update, validate in sequence. Triggers on "/pipeline", "/pipeline [run-folder]", "full pipeline", "process everything".
| name | seed |
| description | Initialize a batch processing job for one or more SWARM experiment runs. Sets up the queue, calculates claim numbering, creates the extract task. The entry point for pipeline processing of new runs. Triggers on "/seed", "/seed [run-folder]", "start processing", "queue this run". |
| user-invocable | true |
| allowed-tools | Read, Write, Glob, Bash |
| context | fork |
Target: $ARGUMENTS
Initialize pipeline processing for SWARM experiment runs.
Claim numbers are global and never reused:
# Find highest existing claim number
highest=$(ls vault/ops/queue/*.md vault/ops/queue/archive/*.md 2>/dev/null | \
grep -oE '[0-9]+\.md$' | grep -oE '[0-9]+' | sort -n | tail -1)
next_start=$((highest + 1))
echo "next_claim_start: $next_start"
---
type: extract
run_id: [run folder name]
run_path: runs/[run-folder]/
next_claim_start: NNN
status: pending
created: YYYY-MM-DD
---
# Extract task: [run-folder]
Run: [[runs/[run-folder]]]
## Run metadata
Type: [single | sweep | redteam | study | calibration]
[Key parameters from run.yaml or summary.json]
## Notes for extraction
[Any flags for the extractor: what to look for, known findings, context]
After creating the task file, add to queue.json:
{
"id": "extract-[run-folder]",
"type": "extract",
"status": "pending",
"target": "runs/[run-folder]/",
"file": "[run-folder]-extract.md",
"created": "[ISO timestamp]",
"current_phase": "extract",
"completed_phases": [],
"next_claim_start": NNN
}