GPT-6 Astra Prompting: OpenAI's Slop Blocklist and Action Rules

OpenAI’s prompting guidance for GPT-6 Astra does something most AI documentation avoids: it names the problem phrases directly. The blocklist targets output patterns that developers have complained about for years, now formalized as things to actively suppress at the prompt level.

Why This Guidance Exists

GPT-6 Astra has known behavioral quirks that differ from previous models: it asks clarifying questions more frequently than predecessors, sub-agents delegate work less than expected, and test suites tend to expand disproportionately when the underlying coding change is small. OpenAI’s documentation addresses these directly rather than leaving developers to discover workarounds through trial and error — a meaningful shift in how the company is supporting production deployments of Astra.

The guidance also reflects how the model was designed to operate: as a concrete-results-first system that should produce something reviewable before seeking approval, not a question-answer system that clarifies before acting. Getting that behavior consistently requires explicit prompting.

The Slop Blocklist

OpenAI identifies these phrases as weakening output — developers are advised to include explicit instructions against them:

• Hedging openers: “It’s worth noting”, “What’s important is”, “In short:”
• Filler intensifiers: “really”, “truly” as standalone amplifiers
• Structural clichés: “Conclusion:”, “Question? Answer” formatting, “The simplest mental model is:”
• AI-isms: “delve into”, “leverage”, “promote” in analytical contexts
• Contrastive framing: “X, not Y” phrasing used as a substitute for specifics
• Invented jargon: made-up hyphenated compounds like “exact-head checks”

The instruction pattern is direct: prompt the model to use clear, concise paragraphs with active voice and simple language. Avoid nested lists and vague qualifiers. The model should write like a practitioner, not a consultant.

The Action Bias Rules

The more operationally significant guidance covers how Astra should interpret instructions. The documentation specifies that phrases like “can you” should trigger action, not clarification. When a user says “can you refactor this function,” the model should refactor — not ask what kind of refactoring is wanted.

The framing OpenAI uses: “the user should be approving a concrete, reviewable result.” The implication is that unsolicited warnings, safety checklists based on hypothetical risks, and clarifying questions before output are all failure modes the prompt should explicitly suppress.

When a model pauses unexpectedly on an agent task, the debugging instruction is specific: force the model to “name and link to the exact SKILL.md file you read, quote the relevant instruction, and briefly explain how it applies.” This turns the pause from an opaque block into a traceable decision.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.