OpenAI shares GPT-6 Astra prompting tips, slop word blocklist

OpenAI's documentation for GPT-6 Astra reveals a model that is more collaborative but also more cautious, sometimes stopping for clarification when users expect it to proceed. To address this, the company provides specific prompt language that tells the model to infer intent and act, while waiting to ask for approval until it has a concrete result. This reflects a design choice to make the model a better teammate rather than a pushover, but it requires developers to adjust their prompts.

The guidance also addresses the model's writing style, which tends toward lists and tables. OpenAI suggests instructing the model to use plain language and avoid 'slop words' — overused AI phrases like 'delve into' and 'leverage' — to produce more natural prose. This is part of a broader effort to make AI output less formulaic and more human.

For development teams, the documentation offers practical debugging tools, such as a prompt that forces the model to identify which skill file caused an unexpected pause. It also warns that GPT-6 Astra can over-test small changes, and advises telling it to rerun tests only when necessary. These tips aim to help developers integrate the model more smoothly into their workflows.