一段提示词让 Opus 5.5 + Jev 自建子代理团队
粘贴即生成 .claude/agents 下的研究员、编码、测试、审查四个子代理,Jev 在每步做路由、上下文、工具闸门和完工判定四类决策,并写出 AGENTS.md 存档。

this prompt is a treasure chest. give it to your Opus 5.5 + Jev stack and watch what happens one paste and your agent builds its own team: -> Opus 5.5 becomes the lead. it plans, writes, and is the only one talking to you -> up to 4 subagents get their own files in .claude/agents: researcher, coder, tester, reviewer -> Jev plugs in as the decision layer between every step -> it writes an AGENTS.md so the whole setup survives the next session the 4 calls Jev makes on every task: > route: which subagent gets the job, or back to the lead > context: which parts of the history each subagent sees. labeled, not compacted > gate: every tool call checked before it runs. irreversible, touches secrets, leaves the repo > done: is the work actually finished, or does it loop once more and the rules that keep it honest: > thresholds live in your config, not in the LLM's head > never simulates a connection, never invents a Jev answer > the destructive-command test runs as a dry run only nothing gets marked verified until it's tested Opus 5.5 does the thinking. Jev makes the hundred small calls in between save it, paste it, let the team build itself
这条是引用着另一条写的:x.com/n01ennn/status/2103542021071978601