Stop treating GPT-6 Astra like its predecessors. OpenAI’s Eric Provencher warns that the new model chokes on verbose instructions, urging developers to strip away the bloat. Overly long skill descriptions and rigid approval rules don’t help; they hinder. When context windows fill with unnecessary text, Astra truncates critical details or stops working prematurely. It’s a classic case of over-engineering the prompt for a model that thrives on brevity.
Skills, stored as Markdown files, are particularly prone to this clutter. If your skill descriptions are vague or broad, Codex struggles to pick the right tool, often loading irrelevant scripts into the context. Provencher advises tight scoping. A Postgres migration skill should only trigger for specific schema changes, not general database queries. Every extra read consumes context, pushing the model closer to summarization limits. Keep it lean. Point to supplementary docs only when necessary.
The same logic applies to repository rules in AGENTS.md. Forcing Astra to read architecture, database, and deployment docs before fixing a typo is wasteful. The model can infer what it needs. Selective referencing works better. Allow local tests on throwaway data without repeated confirmation requests. Explicit permissions for safe operations cut down on friction. Astra doesn’t need hand-holding for routine tasks.
Perhaps most critically, revisit your guardrails. Strict approval workflows designed for older, less reliable models can backfire with Astra. The model interprets literal restrictions too strictly, halting progress when you want it to continue. Define “done” clearly. If you want implementation, testing, and error fixing, say so. Otherwise, Astra might stop after the first step. This shift requires trust, but the efficiency gains are worth the risk. Are your prompts ready for a smarter agent?
