Configure my Claude Code workspace for autonomous agent teams:

In ~/.claude/teams, create the missing roles. Workers run on Sonnet (medium effort) in separate git worktree branches. The adversary runs on Fable.
In ~/.claude/settings.json, set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 and lock the main session to effortLevel: high.
Initialize tasks.json with file-level locking so workers can coordinate directly.
In CLAUDE.md, require an adversary review before any PR.

Show me the config diffs before applying anything.

New favourite use case for opus 5.5:

Give them the whole repo and ask:

“Tell me what we should actually refactor.”

• an orchestrator maps the architecture and conventions
• it sends fresh read-only agents through every subsystem
• it validates, dedupes, and ranks the findings

The result: real code smells, dead code, duplication, and architecture issues instead of unnecessary AI nitpicks.

Prompt 👇

This Grok Bot prompt is literally $300/hr chief of staff https://t.co/dxLaVtmXVJ

ok this one is stupidly good. try it on 5.6 sol extra high in codex!

it plays straight into codex's new strengths, parallel subagents and computer use, and lets 5.6 sol carry a goal all the way to done.

the prompt 👇

I have one complete task to push end to end: [your spec doc, or the full task in detail]. treat it as one end-to-end goal: 𝘁𝗮𝗸𝗲 𝘁𝗵𝗲 𝘄𝗵𝗼𝗹𝗲 𝗽𝗹𝗮𝗻 𝘁𝗼 𝗱𝗼𝗻𝗲, until architecture, implementation, tests, review and the final result all clear the bar. break the goal into independent parts, spawn as many parallel agents as it takes, and give each one its own goal with a clear deliverable, a verification method, and done criteria.

then run it like an operator: schedule the agents in parallel, track progress, synthesize what comes back, resolve conflicts, keep implementing, verify live after every important step, review when it matters, commit when it's ready, and close with a final summary. verification must cover the real end-to-end path: browser and computer use, clicks, keyboard, whatever it takes. don't stop at partial progress unless you're blocked by missing credentials, destructive ambiguity, or conflicting requirements.

做了一个有趣的实验，/grill-me （拷问我），如果这个“我”是AI 会怎么样？

也就是用A harness来拷贝我，B  harness来代替人类回答问题。

herdr 真的是伟大的发明，天生支持Harness之间互相调用。

提示词：

1.  /herdr 开右边pane，启动grok
2. read grill-me skill, understand it. 你扮演人类回答问题，在右侧使用/grill-me skill，我们的目标是clone https://t.co/Wdm7I47qfW

最后问了 16 个问题达成共识，形成文档，效果还不错。😂

当然我感觉可以改进下/grill-me skill，让subagent来回答。