GPT-6 Astra · Creative workflows
GPT-6 Astra Prompting Guide: Keep Tasks Focused and Finish the Job

When Astra turns a small request into a much larger project, the useful correction is often a tighter definition of done. Specify the artifact you want, the decisions it can make, the boundaries it should respect and the checks needed before handing the work back.
OpenAI's behavior guidance describes a model that responds strongly to instructions and may ask clarifying questions or do substantial verification. The prompts below turn those tendencies into a practical workflow. They are starting templates, not a claim that a particular phrase guarantees a better result.
Describe the deliverable before the process
‘Help with our launch’ leaves the agent to infer whether you need research, copy, artwork or implementation. ‘Create one responsive landing page from this approved brief’ gives it a concrete target. Add the intended audience and the business action the page should support so it can resolve routine choices in context.
For the Still lamp example, a useful target is a page with one product image, a short benefits section and an honest next-step control. A full shop, account system and analytics integration are separate work unless the brief includes them.
State the boundary in terms of the work
Name the files or surface that belong to the task when that information is available. Explain which behavior is already approved and which part needs revision. This gives the agent enough context to make a focused change without forbidding necessary fixes.
Avoid blanket commands to skip checks. A useful boundary is ‘verify the changed page at mobile and desktop widths and test its primary control.’ It is less useful to demand either every possible test or no verification at all, regardless of the change.
Set a sensible rule for questions
Questions are valuable when the answer changes the work. If the product price is missing, the agent should not invent one. If the exact spacing between two sections is unspecified, it can usually make a reversible design decision and show the result.
Separate facts from preferences in the brief. Mark unavailable facts as placeholders that must remain visibly unresolved. Give the agent freedom over reversible layout choices while reserving publication, purchases and other consequential actions for the authorization you actually intend.
Implement the approved Still landing-page brief in the existing page. Deliver a responsive preview, not a production release. Keep the current product claims and image. Improve heading hierarchy and the mobile layout. Make reasonable reversible design choices. Ask only when missing information changes the product claim, main action or implementation boundary. Do not invent prices or testimonials. Check desktop and mobile, keyboard access and the primary control; report the result and any unfinished work.Separate creation from acceptance
A deliverable can exist and still fail its purpose. Define acceptance in terms a reviewer can observe. For a campaign brief, every shot needs a subject, duration and motion instruction. For a page, the primary action needs to work. For an image edit, the product identity needs to remain intact.
Ask for a short handoff naming the artifact, checks performed and remaining uncertainty. A long narrative of intermediate work makes it harder to see whether the actual result is ready.
| Vague request | Reviewable requirement |
|---|---|
| Make it professional | Use one type scale, consistent spacing and the supplied palette |
| Finish everything | Deliver these three named artifacts and check each opens |
| Don't ask questions | Resolve reversible layout choices; ask about missing product facts |
| Test thoroughly | Check the changed behavior and its likely dependencies |
Revise the defect rather than restarting the brief
If the desktop page looks good but the mobile hero is crowded, say exactly that. Preserve the approved desktop composition and ask for the mobile issue to be corrected. Repeating the entire creative brief with additional adjectives can unintentionally reopen choices you already settled.
Keep a short list of accepted decisions. On a larger task, that record prevents the next revision from undoing a previous improvement. If a new request conflicts with the accepted direction, make the change explicit so the agent does not have to guess which instruction is current.
The desktop hierarchy is approved. At 390px wide, the headline competes with the product image. Adjust only the mobile hero layout so the headline, image and primary action each have clear space. Preserve desktop spacing and all copy. Show the mobile result and confirm the primary action remains keyboard accessible.Save a workflow only after it earns reuse
A one-off prompt is enough for a one-off task. If you repeatedly prepare the same campaign handoff, save the stable sequence as a skill and keep changing product facts in the individual brief. This avoids turning every request into a growing block of stale instructions.
Review the results before standardizing the process. A skill can make an unhelpful workflow repeatable just as easily as a helpful one. Our companion guide shows a small instruction-only visual-brief skill with a clear input and output contract.
Put your visual brief to work
Once your task contract includes a clear image or video brief, bring that brief into Ciyo to work on the visual assets.