OpenAI shares prompting tips for GPT-6 Astra including a blocklist of slop words
GPT-6 Astra asks clarifying questions more frequently than GPT-5.6 Sol, making it a "more effective collaborator" but sometimes stopping prematurely where users expect continued action OpenAI provides explicit prompting guidance to encourage a "bias towards action," instructing the model to infer user intent and persist until goals are achieved without unsolicited permission requests The model is more sensitive to context and can be derailed by unclear or contradictory instructions in skill file
Analysis
TL;DR
- GPT-6 Astra asks clarifying questions more frequently than GPT-5.6 Sol, making it a "more effective collaborator" but sometimes stopping prematurely where users expect continued action
- OpenAI provides explicit prompting guidance to encourage a "bias towards action," instructing the model to infer user intent and persist until goals are achieved without unsolicited permission requests
- The model is more sensitive to context and can be derailed by unclear or contradictory instructions in skill files like AGENTS.md, requiring developers to audit context documents carefully
- OpenAI has published a "slop words" blocklist (e.g., "delve," "foster," "leverage") and recommends plain-language, active-voice prose over formulaic AI writing patterns
- A debugging prompt is provided to help developers trace unexpected model behavior back to specific skill file instructions by forcing the model to quote exact sources
Why It Matters
This guidance reveals OpenAI's ongoing effort to refine model behavior through prompt engineering rather than solely relying on model updates, signaling that even advanced models require careful instruction tuning for production use. The emphasis on reducing excessive clarifying questions and "slop" language reflects industry-wide concerns about AI output quality and user experience, making this relevant to any practitioner building applications with GPT-6 Astra.
Technical Details
- GPT-6 Astra exhibits increased clarifying-question behavior compared to GPT-5.6 Sol, trading off initiative for collaboration effectiveness; OpenAI recommends prompts that instruct the model to infer intent and demonstrate a "bias towards action"
- The model shows heightened sensitivity to context documents, with unclear or contradictory instructions in skill files (e.g., AGENTS.md) capable of causing work blockage or unexpected direction changes
- OpenAI provides a debugging prompt requiring the model to name exact skill files and quote specific instructions when it pauses or diverges, enabling developers to trace behavior to its source
- A comprehensive "slop words" blocklist includes phrases like "delve into," "leverage," "it's worth noting," contrastive phrasing ("X, not Y"), made-up hyphenated compounds, and summary closers like "In short"
- Default writing guidance emphasizes clear concise paragraphs with active voice, simple vocabulary, and restricted use of lists and Markdown formatting unless structurally necessary
Industry Insight
- Developers integrating GPT-6 Astra should prioritize auditing all context documents and skill files before deployment, as the model's increased sensitivity means previously harmless ambiguities may now cause significant workflow disruptions
- The formalization of "slop word" blocklists by OpenAI suggests a growing industry standard for AI output quality control, and practitioners should adopt similar linguistic guidelines to improve user-facing response quality
- The tension between collaborative clarification and action-oriented behavior represents a fundamental design choice in AI assistants; teams should explicitly configure this trade-off based on their application's needs rather than accepting default model behavior
Disclaimer: The above content is generated by AI and is for reference only.