GPT-6 Astra Prompting Guide

Turn a task contract or an observed Astra failure into a focused prompt change. Keep the requested outcome, authority, and evidence requirements explicit. The skill helps author prompts for Astra regardless of which agent edits them; loading it does not select a model or enable runtime capabilities.

Quick Start

1. Identify the deliverable: new prompt, review, failure diagnosis, migration, or tool description. Preserve the user's exact model target and edit scope.
2. Read the active prompt and relevant loaded instructions, including skills, AGENTS.md, and tool descriptions. Collect the expected result, actual failure when available, authorized actions, and required completion evidence.
3. Use the behavior map below to choose an intervention. Treat a suspected cause as a hypothesis until a trace or representative run supports it.
4. Resolve the conflicting clause at its source within the authorized scope. Avoid layering a second policy over an unresolved first one.
5. Read prompt patterns for wording that fits the failure, or adapt the prompt template for a new prompt. Include only sections the task needs.
6. For runtime changes, read runtime notes. Recheck time-sensitive claims against the official pages in source notes; report unavailable verification.
7. Reuse existing prompt checks or representative traces when available. Compare completion, unnecessary pauses, evidence, writing quality, and verification scope. Report which behavior was actually exercised.
8. Deliver the revised prompt or scoped patch, runtime assumptions, and the reason for each material change. A review request alone returns findings and proposed wording; it does not authorize unrelated repository edits.

Astra Behavior Map

OpenAI's Astra guide identifies these tendencies. They are tuning targets, not guaranteed failures.