OpenAI shares prompting tips for GPT-6 Astra including a blocklist of slop words

Details

GPT-6 Astra asks clarifying questions more often than its predecessor GPT-5.6 Sol — OpenAI frames this as collaborative, but it creates friction in agentic workflows. The official guide recommends explicit "bias towards action" prompts: instruct the model to infer intent from context, treat "can you..." or "help me..." as directives rather than questions, and work through tasks independently until the user's goal is reached before seeking approval. The guide also addresses a structural pain point: GPT-6 Astra is more sensitive to context than previous models, meaning contradictory instructions in AGENTS.md or skill files can cause unexpected stalls or off-task behavior. OpenAI recommends auditing all context documents, giving user-level instructions explicit priority, and embedding a debugging prompt that forces the model to name the exact file and quote the instruction causing it to pause — practical tooling for teams managing complex agentic deployments. On tone and style, OpenAI maintains a formal blocklist of "slop" phrases — delve, foster, leverage, utilize — that GPT-6 Astra overuses by default. The documentation also discourages reflexive list-and-table formatting and unsolicited disclaimers based on hypothetical risks. Prompts should tell the model to drop safety boilerplate unless risks are concrete. For teams migrating from GPT-4o or GPT-5.6 Sol: system prompts may need auditing, as GPT-6 Astra's higher context sensitivity means previously reliable prompts can behave differently without changes to the actual instructions.