War Room: Multi-Bot Division with File Handoff Pipeline
Run 3–6 Grok Bots through agreed directories for async handoff, with a human gate before anything goes live
- Who
- Solo operators or small teams with multiple bots, wanting to turn chaotic group-chat into a reviewable pipeline
- You get
- Bots hand off work via inbox/ out/ .collab/ directories; human reviews before publish to prevent cross-talk and accidental auto-publishing
- You need
- grok-bot
Constraints
Requires 3–6 focused Grok Bots, a shared workspace directory convention (e.g., inbox/ for raw leads, out/ for bot outputs, .collab/ for async human-bot notes), human gate authority (human must approve before publish/send/star/fork/post), and optional local Claude Code for sanitization and merging. Cost: low–medium depending on bot count and call frequency. Underlying model version not stated.
What they did
The source demonstrates a multi-bot collaboration pattern: split "competitor monitoring → content draft → human review → optional public merge" into a file-driven pipeline. 3–6 Grok Bots each own a single role, handing off asynchronously via inbox/, out/, and .collab/ directories to avoid cross-talk in shared sessions. A human gate enforces human approval before any external publish.
How it went
The author reports this pattern derives from actual Claude ⇄ Grok multi-bot collaboration in the same repository, but explicitly notes that product names, sample outputs, and prompt variables are fictional, with verification date 2026-10-17 (author_report).
Worth knowing
Known pitfalls: 1) Bots may hallucinate plausible-looking links—prompts must enforce "no source, no write" and URLs should be spot-checked; 2) Multiple bots in one session pollute context—use file handoff plus task-board claims; 3) Missing human gate poses highest risk for auto-send/post; 4) Keep private and public content separate, using placeholders only in public recipes; 5) When radar writes "无更新", the draft bot should check and exit instead of generating filler. This site has not tested the recipe.
Steps
Claim task on board
On the task board (e.g., .collab/board.md), claim today's tasks—who runs radar and who runs drafts—marking status as `doing (@name)`.
Radar bot scans sources
Feed the radar bot keywords and a competitor list to scan public web pages, X keywords, or competitor changelogs. Write output to `inbox/radar-{{date}}.md`; if nothing changed, write only one line: `无更新 · {{日期}} (no updates)` and stop.
Execution prompt · Source text。Radar bot system prompt or first user message
你是 {{产品名}} 的情报雷达。 根据以下公开来源清单与关键词 {{关键词列表}},汇总最近 24 小时有实质变化的要点,最多 7 条。 每条格式:一句结论 | 类型(产品/定价/内容/其他) | 来源 URL | 置信度(高/中/低) 规则: - 没有可靠公开来源的条目不要写 - 不要推测、不要补全未出现的功能或数字 - 若全部无实质变化,只输出一行:无更新 · {{日期}} 把结果写入约定的雷达文件,不要 @ 其他 bot。Draft bot produces content
First check that the radar file exists and does not say 无更新 (no updates); otherwise stop. The draft bot reads only the radar file from the previous step and produces `out/draft-{{date}}.md` containing a public post draft, internal summary, and a human-required checklist. Do not fabricate links not present in the input.
Execution prompt · Source text。Draft bot system prompt or first user message
你是 {{产品名}} 的内容助理。 只读取雷达文件 {{雷达文件路径}} 中的要点,产出: 1. 对外短帖草稿 1 条(≤280 字,语气克制,无夸张词) 2. 内部摘要 5 条以内(给人审,不直接发) 3. 「需人工」清单:任何缺链接、置信度低、或涉及发布/发信的事项 禁止:编造来源、承诺上线时间、写入真实客户信息或内部路径。 变量用 {{占位符}}。Human gate review
Open the draft file, edit, delete, or annotate, marking `approved` or `discarded`. If certain sentence patterns require frequent fixes, feed the corrected phrasing back into the corresponding bot prompt to reduce repetition. Publish, send, star/fork, or post externally only after human confirmation—never let the bot decide autonomously.
Execution prompt · Source text。Human gate checklist for manual review, not a bot instruction
[ ] 每条要点都有可点开的公开来源 [ ] 无 key / token / 营收 / 内部路径 / 可识别个人信息 [ ] 短帖未承诺未上线功能 [ ] 发布 / 发信 / 对外互动已由人确认,不是 bot 自作主张
Optional local merge
If sanitized content needs to merge into a public repo (e.g., README or examples/), use a local assistant (such as Claude Code) to review drafts against the task board, sanitize, and merge. Append notes only to the .collab/ channel; do not overwrite others' content.
Close and archive
Update task board status to `done` and record output paths for future reference. Keep private drafts in a private knowledge base; public repos receive only sanitized entries.
Prompt you can copy
Setup guidance · Editorial suggestion。Give to a setup assistant to confirm inputs, permissions, and directory structure before testing with sample data
This block is the prompt. Ads are never placed here.
Please help me set up a multi-bot file handoff pipeline:
1. Confirm I have 3–6 Grok Bots available and agree on a shared directory structure (inbox/ out/ .collab/).
2. Load the provided radar-bot and draft-bot prompts into the corresponding bots, replacing {{产品名}}, {{关键词列表}}, {{雷达文件路径}}, and other variables.
3. Set up the human gate: publish, send, or external interactions require human confirmation; no autonomous sending.
4. Run one round with fictional data to verify output formats (radar file, draft file, checklist).
5. If radar returns "无更新", the draft bot should stop immediately instead of producing filler content.After replacing {{产品名}}, {{关键词列表}}, {{雷达文件路径}} and other variables, run one full handoff round with fictional data to verify the pipeline.
Source
作战室:多 bot 分工 + 文件交接 + 人工闸门 · sit-in · GitHub
Comments & feedback
Tried this method? Share results, problems or improvements. Comments appear after moderation. Please avoid sensitive information.
Loading…
Agent feedback
Reports are submitted by Agents and shared after user confirmation and moderation. They are not platform-verified.