菜单
Run a B2B task across three bounded agents (Analyst / SDR / Auditor). The agent in lane produces work; the others refuse and name the correct agent. Built for the AI Agent Economy hackathon — agents that compete on real tasks at agency price points ($500/month).
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
Govern AI spend without surveilling your team — the AI-spend governance loop. Set OKRs, harden the right context (tray→slate), bounded agents surface a finding ready for ratification, ratify a policy to the provenance trail, get a spend-ROI report that cites its sources. The leader governs usage; the team's private context is never read.
The parallel per-agent swarm that beats cold-start. After install, each bounded agent reads its canonical source (git/claude-code → Analyst, granola → SDR, cross-stream → Auditor) in parallel and posts a candidate — the one fact, commitment, or risk worth acting on — before any deliberation. Read-only, LLM-optional. Run this to show the vault won't start empty.
Twelve-agent deliberation on the user's psychological state across four registers (Structural / Somatic / Temporal / Symbolic). Asks three forced-choice questions, runs all 12 agents via Opus 4.7 in parallel, surfaces their disagreement grouped by register, and writes the user's correction — tagged from the canonical taxonomy — to the local vault.
The 60-second taste of Liminal. Runs one bounded deliberation on a sample brief — three agents read the same brief and disagree, at least one refuses out of its lane, you correct one read, and the correction becomes the record. Ends by pointing at the desktop app so the vault outlives the session. Run this first.
Audit an existing Claude Code workspace for canon-substrate gaps — the failure mode where users build sophisticated agent stacks without a CORPUS_MAP / SCHEMA / decisions/ layer underneath, leaving every agent's "consult on demand" reference list resolving to nothing. Reports ship-blocking issues, fragile-but-working issues, and noise. Use when invited into someone's `.claude/agents/` directory to do an audit-and-recommend pass.
Materialize the canon-substrate foundation pack into a target workspace. Reads a diagnose-substrate JSON report + a small config file, copies the recommended template files into the target, substitutes high-frequency placeholders (PROJECT NAME, OWNER NAME, sensitivity tiers), and emits a landing prompt the user pastes into their Claude Code session to finish the customization. Pairs with diagnose-substrate — diagnose first, scaffold second. Step (b) of SPEC §14's productized Diagnostic + Foundation Pack pattern.
| name | agency |
| description | Run a B2B task across three bounded agents (Analyst / SDR / Auditor). The agent in lane produces work; the others refuse and name the correct agent. Built for the AI Agent Economy hackathon — agents that compete on real tasks at agency price points ($500/month). |
| disable-model-invocation | true |
| allowed-tools | Bash(node *) |
| argument-hint | <task description> |
You are running a single B2B task across three bounded agents. Each agent has a specific lane:
The agent in-lane produces real work. The others refuse explicitly and name the correct agent. Refusal is the feature, not an error.
Whatever the user puts after /agency is the task. Examples:
/agency teardown of cofeld.com — competitive moat analysis/agency draft cold email to maya at northstar capital, asking for a 30-min follow-up/agency is the cold-email draft from yesterday ready to send?If the user invoked /agency without an argument, prompt:
"What B2B task? (e.g. teardown, outreach draft, ship/no-ship review)"
Pass the task to the orchestrator:
node ${CLAUDE_PLUGIN_ROOT}/skills/agency/run.js "<task description>"
The orchestrator returns JSON with:
vault_id — deliberation row id in the encrypted vaultanalyst — { interpretation, refused: bool }sdr — { interpretation, refused: bool }auditor — { interpretation, refused: bool }Show all three agent outputs labeled clearly. Lead with the in-lane agent's work.
If one agent refused, show the refusal verbatim — that's the feature beat.
Example output structure:
ANALYST (in lane)
[full structured output]
SDR (refused)
That's the Analyst's lane. I run the move; the Analyst does the homework.
AUDITOR (in lane, judging the Analyst's work)
[verdict + gap or readiness reason]
If the user pushes back on any agent's output (e.g., "the Analyst missed X" or "the Auditor was too soft"), use the existing /check store-correction flow to tag and store:
node ${CLAUDE_PLUGIN_ROOT}/skills/check/store-correction.js <vault_id> "<agent_name>" "<reason>"
Tag from the canonical taxonomy in lib/correction-tags.js.