OpenAI's GPT-6 Astra guide targets AI slop and stalling

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.