抄作业 / 教程 / Grok Bot

做了什么

Sanitized ops lessons for running a Grok Bot fleet beside Claude—hub/lanes, token thrift, dual-runtime source of truth, and room-noise rules.

场景

未归类

资料完整度

• ✓ 作者与原链接有
• ✓ 做了什么有
• ✓ 来源原文摘录有
• — 截图 / 媒体来源未提供
• ✓ 步骤 / 章节有
• — 提示词 / 配置原文来源未提供
• — 结果 / 数字来源未提供
• ✓ 风险提示有
• ✓ 相关条目有

— 表示来源没有给出，本站不补写。

来源原文摘录

Lessons from operating Grok Bot (Orangutan / Chief of Staff) alongside an existing Claude Code fleet — written so others can reuse the patterns. Companion spirit to running-an-ai-fleet, claude-operator-kit, and tiered-agent-memory.

摘自项目 README，照录原文，未翻译。

来源里的步骤 / 要点

1. Chat is disposable; disk is durable. Write decisions to hub/strategy (or equivalent) the moment they lock. Never treat a thread as memory.
2. Hub / lane / worker. One Chief of Staff dispatcher; long-lived specialist lanes; short-lived workers with one report.
3. One owner per repo. Two agents editing the same working tree is a race, not parallelism.
4. Director stays thin. Bulk reads, smash passes, and census work leave the hub. Escalate on verified failure, not vibes.
5. Memory tiers. Tiny always-on profile; project docs when in-project; archive retrieve-on-demand (including a Drive mirror Claude can read).
6. Agents report state changes only (preview URL ready, clear/fail, Jason blocker).
7. No ack-only pings. No FYI. No mid-smash. No multi-topic essays.
8. One user ping per real CLEAR, human blocker, or apex flip.

照录来源原文，未翻译。

提示词

来源没有给出可直接复制的提示词原文。

怎么复制

1. 打开仓库：github.com/jasonpalmer1/grok-fleet-lessons ↗
2. 读 README，把仓库链接发给你的 Bot，让它按 README 复现
3. 先在小范围跑通，再加 routine

风险与限制

• 作者的结果不保证能复现，以原帖为准

用量预估

用金属计算器估一下