GPT-6 Astra Image Prompts: Score the First Result Before You Rewrite

Write the brief as checks

Before you judge an image, write four or five points that can pass or fail: the hero product, where text will go, the palette, what must not appear. Vague goals such as 'make it pop' cannot be scored.

Let a pass be a pass

Our first image, a square feed photo of a honey-lavender latte, was scored against four points. Astra marked all four as passing, gave a reason for each, and recommended no changes. It also suggested testing the real headline in the space before deciding on any revision.

That is a useful outcome. Without the checklist, many people would regenerate a good image out of habit and pay for a worse one.

Name the one point that failed

The second image was a 9:16 story. Our prompt asked for the bottom fifth to stay plain for a line of text, and the model delivered a flat cream band with a hard edge across the frame.

Astra scored it 4 of 5. The failed point was 'one natural photograph': the band read as an added footer, not part of the scene. The text space itself passed.

Score the selected story image against my brief before I post it: (1) the latte is the clear hero, (2) the bottom fifth is clear enough for one line of text, (3) the whole frame reads as one natural photograph, (4) cream, lavender and honey-gold palette, (5) no text or logos. Mark each point pass or fail with the reason. If anything fails, rewrite the prompt to fix only the failed points and keep what passed. Do not generate yet.

Fix the failure with one edit

Astra's rewrite changed only the failed point: replace the disconnected band and its seam with a natural surface, and keep the latte, hand, window, lavender, palette and clear bottom area. Because the rest of the image passed, we applied it as a Quick Edit on the existing image rather than a new generation. The Run button showed 4 credits.

Make this read as one continuous natural photograph. Replace only the flat cream band at the bottom and its hard seam with a cream café tabletop that continues the scene, with believable perspective and the same warm window light. Keep the bottom fifth quiet and clear for one line of text. Keep the latte, hand, window, lavender and palette unchanged. No text, logos, panels or borders.

Score the fix too

The edit replaced the band with a round marble café table that continues the scene, and the latte, hand and window stayed the same. The seam was gone.

One new issue appeared: the tabletop has dappled window shadows, so the bottom area is busier than the flat band. It still has room for text, but you should check contrast with your real line of text, or add a soft solid box behind it.

Use the same loop for every image

Brief as checks, score, change only what failed, score again. The loop is short, and each step leaves a record of why the image changed.

Other creators describe the same discipline for sequences. On September 18, 2026, @dwarffy_ advised keeping the scene description locked and changing only what is new, and treating a sudden change in a frame as prompt drift. That post is about image sequences in another tool; the scoring table here is our own adaptation.

• GPT-6 Astra image generation prompts
• Change one detail in an AI image