OpenAI advises simplifying skills and guardrails for the GPT-6 Astra model

Eric Provencher from OpenAI published recommendations on 12 September 2026 for developers switching to the GPT-6 Astra model, focusing on simplifying prompts, skill files and approval rules (guardrails). According to the company, accumulated instructions that worked with previous models unnecessarily take up model context or cause work to end earlier than the user expects.

According to the article, skills are Markdown files containing instructions that the model selects based on the name and description stored in its context. Too many skills, or unclear and overly general descriptions, force the model to shorten the descriptions, losing information needed to make the right selection. Provencher recommends keeping each skill description short and tied to a specific task (e.g. a database schema migration), and the main skill document should only link to supplementary files that are loaded when needed. Similarly, for AGENTS.md, he recommends avoiding requirements to read all documents (such as architecture.md, database.md, deployment.md) before every change, instead linking to them selectively based on the task type, and explicitly allowing safe operations, such as running tests on local data, so the model does not have to repeatedly ask for confirmation.

According to OpenAI, the GPT-6 Astra model has better judgment than the earlier GPT-5.6 Sol model, but it may interpret overly strict approval rules set for older models literally and stop work prematurely. Provencher therefore recommends reviewing such rules during the transition and explicitly defining what “done" means in the prompt—that the model should not only implement the task, but also run it, check the result and fix errors.

These recommendations follow previously published, more general tips for prompting the GPT-6 Astra model, which, according to the company, covered greater initiative from the model, a list of unwanted expressions (“slop words") and the extent of delegation to sub-agents. The earlier report stated that the model asks more clarifying questions than GPT-5.6 Sol and follows longer instructions better, but is more sensitive to the content of skill files.