OpenAI hands developers a prompt to stop GPT-6 Astra waiting for permission
The model documentation concedes Astra asks for clarification more often than GPT-5.6 Sol and sometimes stops early. Every remedy on the page is more instruction text, loaded into the same context the docs blame for stalls.
The Engineer · Build desk
What happened
- OpenAI's documentation for GPT-6 Astra says the model asks clarifying questions more often than GPT-5.6 Sol rather than assuming, which OpenAI describes as making it a more effective collaborator.
- To restore initiative, OpenAI supplies prompt text telling the model to infer intent and show a bias towards action, reading "can you" or "help me" as instructions to act rather than openings for questions.
- A separate debugging prompt requires the model to name and link the exact SKILL.md it read and quote the instruction that made it pause, ask for permission or diverge from the user's intent.
Why it matters
- What it costs cost Following the guidance means editing every system prompt and auditing each skill file the model can read, then re-establishing which instructions win; that labour sits with whoever owns the prompt tree.
- Who has to decide decision With approval deferred until a reviewable artefact exists, sandbox scope and repository permissions become the real limit on a self-started run, so those settings now need reviewing alongside the prompt.
- What doesn't add up contradiction One block on the page bars hyphenated adjectives while another instructs the model to perform read-only actions, so a model obeying the style rule literally strips vocabulary the permission rule depends on.
Next story · Build · Oct 5
OpenAI opens its 28-day Codex pledge by speeding GPT-6 Astra and Sol to about 50 tokens a secondConfirmed2 publishers
The bias-towards-action block is an allowlist with a deferred gate. OpenAI's recommended text names what the model may do unsupervised: create isolated work trees or checkouts, resolve merge conflicts, perform read-only actions, open draft PRs, with a carve-out for anything clearly destructive or irreversible [4]. Approval arrives afterwards. The docs put it as a rule about what the human is handed: the user should be approving a concrete, reviewable result [5]. OpenAI also recommends stripping unsolicited warnings, disclaimers and hypothetical-risk checklists out of prompts [11].
So the gate moves from before the work to after it [17]. If a run now produces a branch and a draft PR before anyone is asked, the thing bounding the blast radius is repository permissions and whatever sandbox the agent runs in, and the prompt is just a request.
Here is the part worth pricing. The same documentation says Astra follows longer instructions better than its predecessors but is more sensitive to context [6], and that unclear or contradictory lines in skill files such as AGENTS.md can make it block work or veer off unexpectedly [7]. The prescribed fixes are three further blocks of instruction text: initiative, prose style, and the slop-word list [16]. The cure is measured in the same units as the failure. Hence the accompanying advice to audit every skill file and context document the model can reach and to give user instructions explicit priority [8].
The diagnostic prompt is the strongest engineering on the page, because it makes the model produce evidence: name and link the exact SKILL.md it read, quote the instruction, and separate an explicit skill requirement from its own interpretation of a guideline [9]. That converts "it stopped" into a path and a line you can go and delete.
The blocklist is harder to operationalise than it looks. It covers literal strings such as "delve into", "it's worth noting" and "Conclusion:" in conclusions, but at least five of its entries are sentence shapes rather than tokens: "Question? Answer", "This isn't about X. It's about Y", contrastive "X, not Y", compound descriptions and hyphenated adjectives, and any mention of what the model will not do [14][15][18]. Substring linting catches the first group and nothing in the second. The presence of "Conclusion:" on a list of things to avoid in conclusions reads like documentation written by someone who has graded a great deal of model output. And note that the ban on hyphenated adjectives sits on the same page as an initiative prompt instructing the model to perform read-only actions [19]; applied literally to technical prose, that rule removes a compound carrying a permission boundary.
What the source does not contain is a number. There is no measured stall rate, no before-and-after on Sol, nothing that would let you estimate how much of your own agent's idling this prompt text recovers [10]. For these blocks to transfer, your stalls have to be the over-clarification the docs describe [1][2] rather than a contradictory skill file [7], and those have opposite fixes: one wants more instruction, the other wants a line deleted. That argues for running the SKILL.md diagnostic first and pasting the initiative block second.
What to watch
- Whether OpenAI publishes a measured stall or over-clarification rate for Astra rather than prompt text alone.
- Whether these initiative and style blocks ship as product defaults, which would take them out of every developer's prompt diff.
- Whether the diagnostic prompt still traces behaviour in stacks whose context files are not named SKILL.md or AGENTS.md.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence62
- Adoption
- Not scored
- Hype gap+10
- Incentives50
- Confidence65
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
According to OpenAI, GPT-6 Astra asks clarifying questions more often than GPT-5.6 Sol instead of making assumptions on its own, which OpenAI says makes it a "more effective collaborator".
- [2]
The trade-off OpenAI states is that the model sometimes stops where users expect it to keep going.
- [3]
To push the model toward more initiative, OpenAI recommends a prompt telling it to infer the user's intent and scope from context and show a "bias towards action"; phrases like "can you...", "I want to..." or "help me..." should be treated as calls to act rather than invitations for follow-up questions.
- [4]
The recommended prompt tells the model to work independently toward the user's goal, giving as examples creating isolated work trees or checkouts, resolving merge conflicts, performing read-only actions and creating draft PRs, unless the actions are clearly destructive or irreversible.
- [5]
OpenAI's guidance says the model should wait to ask for approval until it has already prepared a concrete, reviewable result, prompting it with the line "The user should be approving a concrete, reviewable result."
- [6]
OpenAI says GPT-6 Astra follows longer instructions better than its predecessors but is also more sensitive to context.
- [7]
Unclear or contradictory instructions in skill files such as AGENTS.md can cause GPT-6 Astra to block work or veer off unexpectedly, according to OpenAI's documentation.
- [8]
OpenAI recommends auditing all skill files and context documents the model can access, and giving user instructions explicit priority.
- [9]
OpenAI recommends a debugging prompt that requires the model to name and link the exact SKILL.md file it read, quote the relevant instruction that made it ask for permission, pause, leave work unfinished or diverge, explain how it applies, and distinguish explicit skill requirements from its own interpretation of guidelines.
- [10]
OpenAI's model documentation for GPT-6 Astra spells out where the model tends toward unwanted behavior and how developers can work around it.
- [11]
OpenAI says unsolicited warnings, disclaimers or safety checklists based on hypothetical risks should be dropped from prompts.
- [12]
GPT-6 Astra tends to structure responses with lists, tables and Markdown formatting, and reuses the same phrases across sessions.
- [13]
OpenAI's style prompt tells the model to default to clear, concise paragraphs each developing a single main idea, to use lists only when information is truly parallel, sequential or more easily comparable, to avoid nested lists, and to favour the active voice.
- [14]
OpenAI calls typical AI phrases "slop words" and lists items to avoid including "Conclusion:" in conclusions, "delve into", "promote", "use/leverage", "it's worth noting", "what's important is", "Question? Answer", "This isn't about X. It's about Y", "really/truly", compound descriptions and hyphenated adjectives, summary closers such as "In short:..." and "The simplest mental model is:...", and contrastive phrasing such as "X, not Y".
- [15]
OpenAI's guidance also says made-up hyphenated compounds such as "exact-head checks" or "editorial-row layouts" should be avoided, and that the model should state what it is doing rather than mentioning what it will not do, what remains unchanged, or how it will categorise results.
- [16]
The documentation's remedies add three blocks of instruction text to the model's context: an initiative block, a prose-style block and the slop-word blocklist, all loaded into the same context the docs say can make the model block work.
- [17]
Under OpenAI's recommended prompting, the approval gate moves from before the work to after it: the model performs the listed bounded actions on its own and asks for approval once a reviewable artefact exists.
- [18]
At least five entries on OpenAI's slop-word list are sentence shapes rather than fixed strings, so a substring filter cannot implement the list.
- [19]
The same OpenAI documentation bans hyphenated adjectives in the model's output while its recommended initiative prompt instructs the model to perform "read-only actions", a hyphenated compound.
Sources
2 independent publishers whose own reporting we read for this story.
- the-decoder.comOpenAI shares prompting tips for GPT-6 Astra including a blocklist of slop words
1 article · September 5, 2026
- theneuron.aiAI Skill of the Day Digest — GPT-6 Astra: 7 Better Ways to Work With It
1 article · September 6, 2026
Topics and entities
Topics
- Prompt Engineering
- Agent autonomy and approval controls
- AI Coding Agents