Rumored 12-page Altman guide on building a GPT-6 Astra project team
A 12-step blueprint said to have leaked from OpenAI on turning a single brief into an Astra agent team that researches, builds, reviews and delivers work.
Sam Altman, CEO of OpenAI, just released a 12-page PDF on how to turn GPT-6 Astra into an entire project team It is more useful than most paid AI courses: This is a 12-step blueprint for turning one brief into a team that researches, builds, reviews, fixes, and delivers work with proof. step 1 → write the finish line first: exact deliverable, constraints, approval boundary, and the check that proves the project is done step 2 → put one project manager in charge of the board: every job gets an owner, a state, a next action, and a reason when it is blocked step 3 → split the brief into job cards with one output and one acceptance test each; a team built from vague roles just produces more vague answers step 4 → recruit the smallest useful roster: manager, researcher, builder, reviewer; add a specialist only when it owns a real artifact step 5 → give each role a case file, not the entire conversation: current state, relevant sources, allowed files, open questions, and the exact check step 6 → make tools a contract: what each role can read, what it may change, which receipt it returns, and where it must stop step 7 → run independent jobs in parallel, but never let two workers edit the same source of truth; merge only after their local checks pass step 8 → make review independent: compare the artifact with the original brief, verify claims against sources, and open or render the actual output step 9 → send failures back as precise repair packets: failing assertion, artifact version, allowed fix, and the same verifier for the rerun step 10 → keep publishing, spending, deletion, deployment, and external messages behind a concrete human approval request step 11 → save the team as operating files: role contracts, job-card templates, tool rules, board schema, and verification commands step 12 → test the system on real projects: count accepted deliverables, retries, human interventions, latency, and full model-plus-tool cost the result: one brief becomes a coordinated project with a visible trail from request to finished artifact — and no one has to guess whether it actually worked. Bookmark this, send the operating pattern to your AI agent, and read the full 12-page field guide below ↓
This post quotes another post: “When to use GPT-6 Astra and when Sol is enough”