Anvyll Download

How to prompt GPT-6 Astra

OpenAI Markdown section headers Updated

Prompt GPT-6 Astra with markdown sections (Context, Task, Constraints, Output format), written outcome-first: describe what should be true when the work is done and let the model choose the route. Astra follows instructions more closely than earlier GPT models. That makes it more sensitive to everything in its context, including conflicting instructions in files you forgot were loaded.

What makes Astra different

OpenAI highlights five behaviours to plan around:

  • It asks more questions. Astra collaborates readily, so it may pause for clarification where you expected it to make an assumption and carry on.
  • It follows instructions closely. That includes instructions in skill files and repo docs such as AGENTS.md. Unclear or contradictory guidance in those files can stall it.
  • It formats heavily by default. Answers are detailed and full of lists and tables, with some recurring stock phrases.
  • It delegates cautiously. In multi-agent setups it may hand off less work than you want.
  • It tests thoroughly. On coding tasks it tends to test extensively before calling the work done.

OpenAI also notes it reaches its results with far fewer output tokens than earlier models, so cost per task can drop even at higher per-token pricing.

The prompt structure: markdown sections

Astra reads markdown headers cleanly and doesn’t need XML.

## Context
Next.js 15 app, TypeScript, Tailwind. Checkout uses Stripe Elements.
Mobile users report the pay button sits below the fold on small screens.

## Task
Fix the checkout layout so the pay button is visible without
scrolling on a 375×667 viewport.

## Constraints
- Don't change the desktop layout.
- No new dependencies.

## Output format
A diff of the changed components, then two sentences on the approach.

## Success criteria
- Pay button fully visible at 375×667.
- Desktop screenshots unchanged.

Spend your words on what changes the result: the goal, context that matters, which rules win when they conflict, what the model may assume without asking, what the output looks like and when to stop.

Get it to act, not ask

If Astra keeps pausing for approval, state your expectations explicitly. It should work out what you intend from the conversation, lean towards action, and carry the task through to completion.

Two more lines help:

  • Treat phrases like “can you…” or “I’d like to…” as instructions to do the work, not as questions.
  • Before asking a clarifying question, finish whatever work is already clearly authorised, so the question is about something concrete.

Use direct verbs: “implement”, “fix”, “write”. Asking for “a suggestion” or “a plan” is exactly what you’ll get.

Say which instructions win

Because Astra weighs its context so heavily, spell out priority. For example: explicit user instructions override guidance in skill files.

When debugging why it stopped, you can ask it to quote the exact instruction that made it pause and name the file it came from.

Give load-bearing context only: the stack, the audience, named files, existing patterns, decisions already made. Don’t invent file or component names to make the prompt look complete.

Turn down the formatting

If you want prose, ask for it directly: clear paragraphs, each with one idea, and lists only for genuinely parallel or sequential items. You can also name the stock phrases you don’t want to see.

In the output section, state the exact deliverable. For code, ask for diffs instead of whole files and cap the explanation. For messages, name the parts, such as a subject line.

Constraints and testing

Put hard limits in their own section rather than scattering them through the task. Phrase them positively where you can. Name the files in play, what must keep working, and whether new dependencies are allowed.

Astra’s thorough testing is usually good, but it can be excessive for small, reversible changes. Calibrate it rather than switching it off: ask for tests only where they meaningfully verify the change.

Reasoning effort and delegation

  • Astra has no none reasoning effort. If you migrate from none or minimal, start at low. Otherwise keep your existing effort and re-measure.
  • To increase delegation, tell it to hand off parallelisable work whenever that saves time or improves quality.
  • For demanding tasks, finish with two or three observable success criteria rather than a list of process steps.

Migrating from GPT-5-series prompts

  1. Add the act-don’t-ask guidance if Astra pauses more than you want.
  2. Add an explicit prose instruction if you don’t want list-and-table output.
  3. Audit skill files and AGENTS.md for vague or conflicting rules. Astra now takes them seriously.
  4. Check OpenAI’s model page for sampling parameters and API changes before switching the model field.

Sources