OpenAI's GPT-6 Astra guide targets AI slop and stalling
OpenAI's guide for GPT-6 Astra lists phrases the model should drop, and shows how to stop it pausing mid-task or over-testing code.
Symbolic image: hands typing at a developer workstation while two monitors glow with abstract text blocks and a status bar.
OpenAI's prompting guide for GPT-6 Astra works around three habits developers keep hitting: the model pausing to ask, its stock AI phrasing, and runaway test runs on small code changes.
At a glance
- GPT-6 Astra asks clarifying questions more often than GPT-5.6 Sol, which OpenAI frames as collaboration but users may read as stalling.
- To restore momentum, OpenAI suggests prompting for a tendency to act, with approval requested only after a concrete, reviewable result.
- Contradictory skill files such as AGENTS.md confuse the model; OpenAI advises auditing them and giving user instructions priority.
- Slop terms to block include ‘delve into,’ ‘leverage,’ ‘foster,’ ‘promote,’ ‘Conclusion:,’ ‘In short:,’ and the ‘X, not Y’ pattern.
- On coding tasks the model over-tests before finishing; OpenAI recommends re-running tests only for new or unresolved failures.
OpenAI shipped GPT-6 Astra with a detailed prompting guide, and much of it reads as a set of workarounds for how the model behaves by default. Three habits get the most attention: it stops to ask, it writes in stock AI cadence, and it tests small code changes far harder than the change warrants. The-decoder covered the guide on September 5, 2026.
Getting the model to keep going
Astra asks more questions than GPT-5.6 Sol instead of making its own assumptions. OpenAI calls that a more effective collaborator, while acknowledging the flip side: the model halts at points where the user expected it to carry on.
The suggested fix is a prompt that tells the model to infer intent and task scope from the instructions and the conversation so far, and to show a tendency to act and follow through until the task is done. Phrasings such as ‘can you’ or ‘help me’ should be read as work orders, not as questions. The carve-out is for steps that are clearly destructive or irreversible.
On approvals, the order matters: produce a concrete, reviewable result first, then ask. Unprompted warnings and safety checklists should be left out.
Skill files are now load-bearing
Astra follows longer instructions more reliably, but it is also more sensitive to whatever context you hand it. Vague or conflicting rules in skill files such as AGENTS.md therefore land harder than they used to. OpenAI's advice is to audit those files for contradictions and to state plainly that user instructions win.
For debugging there is a concrete move: make the model name and link the exact skill file, and quote the line that made it pause or change direction. That converts a vague complaint about the model ‘just stopping’ into something you can open and edit.
The slop blocklist
Left alone, the model reaches for lists, tables and Markdown, and recycles the same connective phrases. For clean prose OpenAI asks for concise paragraphs that each develop one idea, plain language with familiar words and precise verbs, and active voice with direct statements.
Discouraged patterns include:
- ‘Conclusion:’ and wrap-ups like ‘In short:’
- ‘delve into,’ ‘leverage,’ ‘foster,’ ‘promote’
- ‘it is worth noting,’ ‘importantly’
- rhetorical question-then-answer beats
- contrastive framing such as ‘this isn't about X, it's about Y’
- invented hyphenated compounds
One rule goes past style. The model should state the action it intends to take rather than listing what it will not do. Jargon is allowed only where it genuinely aids understanding.
Sub-agents and test sprawl
OpenAI names two further rough edges. Astra delegates to sub-agents less often than intended, so developers should spell out when delegation should happen and how much of the work it should cover. Messages passed between agents can carry grammar or spacing errors.
The second is testing. Before calling a coding task finished, the model verifies exhaustively, which turns a one-line edit into a disproportionate test run. OpenAI suggests calibrating that behavior in the prompt and limiting repeat runs to new failures or problems that remain unresolved.
What we did not verify
The full prompt templates sit in OpenAI's model documentation, according to the report. We did not open that documentation ourselves; everything above rests on the coverage cited below.
A note on sourcing: the two references are the German and English editions of the same newsroom. No independent second source was available at the time of publication.
FAQ
Which words does OpenAI tell GPT-6 Astra to avoid?
The guide flags stock AI phrasing such as ‘delve into,’ ‘leverage,’ ‘foster,’ ‘promote,’ ‘it is worth noting,’ openers like ‘Conclusion:’ and ‘In short:,’ contrastive ‘X, not Y’ framing, and made-up hyphenated compounds.
Why does GPT-6 Astra stop and ask instead of finishing the task?
Astra asks clarifying questions more often than GPT-5.6 Sol rather than assuming. OpenAI suggests prompting it to infer intent from context and act until the task is complete, requesting approval only after a concrete result, except for destructive or irreversible steps.
How do I stop GPT-6 Astra from over-testing my code?
Calibrate the behavior explicitly in your prompt. OpenAI recommends telling the model to re-run tests only when new failures appear or an issue is still unresolved, which keeps verification proportional to the size of the change.