GPT-6 Astra Best Practices: The Complete Prompting Guide

OpenAI's official GPT-6 Astra best practices: initiative, instruction hierarchy, writing style, subagent delegation, testing calibration, plus a free prompts PDF.

GPT-6 Astra is generating the kind of buzz that usually means two things: a flood of secondhand takes, and an official best-practices page that most people never actually read. This guide skips the secondhand part. Everything below is pulled directly from OpenAI's own GPT-6 Astra documentation, organized so you can apply it to your system prompts today.

What GPT-6 Astra actually is

GPT-6 Astra is OpenAI's most advanced model, built for computer use, browsing, software engineering, and professional workflows. OpenAI positions it as showing stronger alignment, more transparent reasoning, and better respect for task boundaries than earlier models — which is also why several of the best practices below are about being explicit rather than assuming the model will infer intent the way older models did.

Initiative and follow-through

Astra responds well to prompts that explicitly grant it room to act. OpenAI's own guidance is to use language that establishes a bias toward completing the task, not just proposing a plan:

"bias towards action and carry the user's intended task to completion"

The practical implication is to treat a capability request as authorization to do the work, rather than gating every step behind a confirmation. If you've been writing prompts that ask the model to "propose an approach and wait," Astra's default behavior expects the opposite framing — describe the intended outcome and let it carry the task through, reserving explicit approval gates for genuinely irreversible actions.

Instruction following and the skill hierarchy

When a project uses skills or tool configurations that carry their own guidelines, Astra needs an explicit statement of precedence, or it can end up treating a skill's defaults as equal to your instructions. OpenAI's documented line for this is direct:

"The user's instructions take precedence over guidelines provided in a skill."

Pair that hierarchy statement with a request for transparency: ask Astra to flag it when a skill's guidance conflicts with what you asked for, instead of silently picking one. That turns a potential silent failure into a visible one you can correct.

Personality and writing style

For prose-heavy output — documentation, explanations, summaries — Astra's default can lean toward nested bullet lists even when the content reads better as paragraphs. OpenAI's guidance addresses this directly:

"Default to using clear, concise paragraphs" and avoid "nested lists unless hierarchy cannot be expressed clearly in prose"

The same guidance calls out a specific tic worth banning explicitly: contrastive framing like "X, not Y." OpenAI's phrasing is to avoid "contrastive framing such as 'X, not Y.'" If your Astra output keeps reaching for that construction, it's worth adding this line verbatim to your system prompt rather than fighting it prompt-by-prompt.

Subagent delegation

Astra's async tool calling makes it genuinely useful at splitting work across subagents instead of running everything in one linear thread. To get that behavior reliably, OpenAI recommends stating it as a default rather than an option:

"If at any point you can parallelize work by delegating tasks to another agent, you should do so."

This works best when it's paired with async tool calling and mid-turn steering — Astra's two headline capabilities for this release. Async tool calling lets delegated work continue executing while Astra keeps reasoning on the rest of the task; mid-turn steering lets you redirect it over WebSocket without losing the work already completed.

Testing and verification, calibrated to blast radius

Astra can over-test by default, adding coverage for changes that don't need it. OpenAI's recommended calibration is to scope test-writing to the actual risk of the change:

"Do not write tests for reversible, low-impact changes that mirror the implementation."

In practice, that means reserving explicit test requests for changes with real blast radius — shared interfaces, migrations, anything security-sensitive — and letting reversible, local edits skip the ceremony. Combined with Astra's misalignment monitoring (built-in safeguards with asynchronous oversight), this keeps verification proportional instead of uniformly maximal.

Where to go from here

If you're migrating existing prompts to Astra rather than starting fresh, the parameter and API changes are a separate concern from the prompting style above — see our GPT-6 Astra migration guide for the exact changes (model name, prompt cache TTL, dropped sampling parameters, the Responses API). For where these capabilities point in terms of actual workflows, see GPT-6 Astra best use cases.