OpenAI has published prompting guidance for GPT-6 Astra, and most of it is instruction on how to switch off behavior the model ships with. Astra asks clarifying questions where GPT-5.6 Sol would have made an assumption. It stops work at the point users expect it to continue. It turns plain answers into lists and tables, repeats the same constructions across sessions, and runs elaborate test suites on trivial code changes. For each of these, OpenAI's remedy is a paragraph of system prompt that the developer has to write.
The company frames the questioning habit as a feature — a model that checks scope is a better partner — while acknowledging the cost, which is that Astra sometimes halts mid-task. The suggested correction tells the model to read intent from context and lean toward acting:
You should infer the user's intent and the scope of the task from the instructions and the conversation context so far. Your task is to demonstrate a tendency to act and to follow through on the user's intended task until completion. If the user expresses the intention to complete new work or resolve an existing problem, continue working persistently until the user's intended goal is achieved. Work independently toward the user's goal (e.g., create isolated work trees/checkouts, resolve merge conflicts, perform read-only actions, create draft PRs, etc.), unless the actions are clearly destructive or irreversible.
In practice the instruction reclassifies a category of phrasing. "Can you", "I want" and "help me" are to be treated as orders to begin work, not as openings for a follow-up question. Approval waits until there is something concrete to inspect; OpenAI's own phrasing is that the user should approve a specific result that can be reviewed. Unsolicited warnings, caveats and lists of safety measures built on hypothetical risks are to be stripped out of prompts.
Astra follows long instructions better than its predecessors and is, at the same time, more sensitive to context. Unclear or contradictory rules in skill files such as AGENTS.md can make it stop work or change direction without warning. OpenAI's advice is to audit every skill file and context document the model can reach, and to rank user instructions explicitly above the rest.
When the behavior is already strange, the company offers a debugging prompt:
If a skill causes you to ask for permission or confirmation, pause, leave requested work unfinished, or diverge from the user's intent, name and link to the exact SKILL.md file you read, quote the relevant instruction, and briefly explain how it applies. Distinguish explicit skill requirements from your interpretation of guidelines.
That a query of this kind is necessary is the admission worth pausing on. Instruction-following is now strong enough that one ambiguous line in a repository file can stop an agent, and opaque enough that finding the line requires a forensic request to the model that read it.
The longest stretch of the guidance is about prose. Astra defaults to lists, tables and markup, and reuses the same expressions from session to session. OpenAI recommends asking explicitly for connected text — short paragraphs, plain language, active voice — with each paragraph developing one idea, lists reserved for material that is genuinely parallel, sequential or easy to compare, and nested lists avoided whenever plain text can carry the hierarchy. The remaining advice reads like a style manual: simple and direct phrasing, familiar words, concrete examples, precise verbs, active voice and direct statements, the main point first with explanation after, sentences that follow from one another, enough support on important points to make the text useful, and professional terminology only where it earns its place.
Then comes the stop-list, which OpenAI calls filler. It covers "Conclusion:" at the head of conclusions, "delve into", "facilitate", "utilize", "it's worth noting", "what matters is", the "Question? Answer" construction, "It's not about X. It's about Y", "really" and "truly", and compound hyphenated descriptors — the examples offered are "exact-match header checks" and "editorial-style row layouts". The model is told to state what it does rather than enumerate what it will not do. Closing formulas such as "In short:…" and "The simplest mental model:…" are out. So are separate announcements about what stays unchanged, what the model declines to do, and how it intends to split and classify its results. Contrasts of the form "X, not Y" and "X is not Y" are to be avoided unless the user asked for the comparison.
A vendor publishing a banned-words list for its own model is an unusual document. Every entry on it is a tic that OpenAI's training produced and OpenAI's writers now find conspicuous enough to name in public. Taken together the section is a description of what GPT-6 Astra sounds like when nobody corrects it, and the correction is billed to the customer — as tokens in a system prompt, on every request.
Two smaller notes close the guidance. Astra can hand work to subagents running in parallel but does so less often than developers expect, so instructions should state when delegation is required and how much work to pass along. Messages between agents, OpenAI adds, may contain errors in grammar and spacing. On coding tasks the model runs detailed tests before finishing, which on small changes produces verification out of proportion to the work; the recommendation is to rerun tests only when new failures appear or problems remain unresolved.
The grammar note is presented as housekeeping and is the strangest line in the document. Put plainly: the text the model produces for other agents is not held to the standard of the text it produces for people, and that is disclosed as a quirk rather than a defect. Anyone building multi-agent pipelines on Astra is being told, in a subordinate clause, that the messages moving between stages are lower-quality output.
More detailed versions of these prompts sit on the GPT-6 Astra documentation page, and there is a migration path for developers who would rather not write them by hand: Codex with the OpenAI Docs skill, invoked as $openai-docs migrate this project to GPT-6 Astra, applies the recommended changes automatically. A model release that ships with its own migration command is not a drop-in replacement. It is a dependency with a breaking change, and the breaking change is the model's disposition.