网传 Altman 分享的 Dots + Codex 多模型分工提示词

Mika @mika_systems原推发布于 2026.10.04

在 X 上打开原推

Sam Altman (CEO of OpenAI) just shared the prompt behind his Dots agent setup put GPT-6 Astra in charge of the approach. give GPT-6.1 Sol a clear implementation task and checks it must pass use Dots to review the handoffs: what Codex should change, why an attempt failed, and whether the finished patch actually solves the task bring Astra into three decisions: → before a large change → challenge the plan and identify what could break → after a repeated failure → compare both attempts before another edit → before "done" → inspect the diff, test results and unresolved risks keep GPT-6.1 Sol on the code. delegate file searches and documentation checks to separate Codex agents, then feed their findings into the implementation the goal: stop carrying plans, errors and test results between chats yourself. give the dot responsibility for reviewing them and sending Codex the next instruction paste this setup prompt into your dot’s chat ↓ "Set up this Dots + Codex workflow for my current project: 1. Have Codex inspect ~/.codex/agents and .codex/agents for explorer, researcher and worker roles. Read the applicable AGENTS.md instructions. Reuse compatible agents; list conflicting instructions and pinned models before proposing replacements. 2. Draft valid TOML files only for missing roles, including name, description and developer_instructions: > explorer: gpt-6-luna, medium effort → trace relevant files, callers and existing tests; no edits > researcher: gpt-6-luna, medium effort → check documentation and API constraints; no edits > worker: gpt-6.1-sol, medium effort → make scoped changes and run targeted tests 3. Draft model = 'gpt-6.1-sol' and model_reasoning_effort = 'high' in ~/.codex/config.toml. Check overrides (project config, active profiles, shell aliases, launch flags, task settings, agents.default_subagent_model and agents.default_subagent_reasoning_effort). Report conflicts without changing them. 4. Draft one workflow rule in AGENTS.md: delegate independent investigations to explorer and researcher, give worker their findings, and return checkpoints before major implementation, after a repeated failure and before completion. Review those checkpoints yourself, then send the same Codex task a revised plan, a targeted investigation or remaining corrections. Show all proposed changes as diffs first. Apply nothing until I approve. After approval, verify the roles load with the intended models. Preserve existing permissions; do not merge or deploy." if you use Codex, steal this Dots setup before everyone else does ↳ https://developers.openai.com/codex/dots

这条是引用着另一条写的：x.com/solomonrojie/status/2105713134174605327