OpenAI suggests that developers optimize prompts for the GPT-6 Astra model by reducing excessive instructions and rigid guardrails. Eric Provencher notes that newer, more capable models require less oversight, as overly broad or conflicting descriptions can consume context and lead to premature task termination.
The company advises keeping skill descriptions in Markdown files short and precise. Developers should link instructions directly to specific tasks and ensure that agents only access necessary documentation. For instance, architecture or database files should be referenced selectively rather than read in their entirety for every minor change.
Provencher emphasizes that strict approval rules, often implemented for older models, may now hinder performance. He recommends explicitly allowing safe, known workflows to prevent the model from seeking unnecessary confirmations. Defining clear completion criteria upfront is also essential to ensure the agent finishes complex tasks.
These recommendations follow OpenAI's recent release of prompting guidelines for GPT-6 Astra. The company encourages teams to regularly review AGENTS.md files and task prompts when transitioning between models to maintain efficiency and avoid redundant processing.
