GPT Image 2.5: Test Sketch and Prompt Sharing
A good image prompt is easy to admire and surprisingly hard to hand to another creator. Does a rough sketch preserve a poster's layout when the person and product change? Can someone reuse a shared prompt without inheriting private source images? The September 8 ChatGPT Images 2.5 update gives teams new controls to test those questions, but an announcement is not a repeatability result.
Create a separate, licensed source portrait with APOB.
This is a reproducible test plan, not a report of two finished Sketch outputs. No APOB integration with GPT Image 2.5 is claimed. A final case study requires the original sketch, two authorized input images, both generated outputs, an actual shared prompt, and notes from a second creator. Without those artifacts, any claim that the layout transfers is still unverified.
Locate the September 8 creation controls.
OpenAI's Images 2.5 announcement describes Sketch, image comments, and prompt sharing in ChatGPT. It separately names GPT-Image-2.5 Flare and Sunburst for API developers. Those names should not be collapsed into a claim that every ChatGPT user selects an API model or that another product has integrated it. Record the exact date, account, interface, and feature availability before testing.
Sketch
The official launch post says Sketch lets a user draw directly in ChatGPT as a visual guide. The current help page documents a mobile-app @Sketch path. Capture the actual entry point on the tested account and the submitted sketch. If unavailable, mark BLOCKED: feature unavailable; a launch screenshot is not a test input.
image comments
OpenAI also describes placing comments on images for focused editing. The proposed test is to annotate one concrete defect after the first output, such as CTA touches the bottom edge; preserve face and headline. Capture the comment and next result. Avoid an untestable instruction like “make it better.” The review should record whether the requested area changed and whether protected elements drifted. That result must come from a real session, not a claim inferred from the feature description.
prompt sharing.
The announcement says users can choose to include the prompt when sharing an image; the ChatGPT release notes also record this control. That is not a guarantee of identical pixels. Inspect what the recipient sees and redact private names, links, images, and campaign details.
Draw a rough layout that can be graded.
The asset is an original vertical creator poster for a fictional, rights-cleared product. The layout has three zones: hook at the top, subject in the center, and CTA at the bottom. Before generating anything, make a one-page grading sheet with zone boundaries, preferred reading order, and protected text. A result can look attractive and still fail this task by moving the CTA into an unsafe crop.
hand-drawn zones
On a 9:16 rectangle, mark a top zone for a short promise, a center zone for the portrait and product, and a lower zone for the action line. Draw rough boxes and label them H, P, and C. The exact pixel coordinates depend on the final canvas, but the sketch must be saved as an image with a version ID. Tell the model that the boxes represent hierarchy, not literal borders or UI widgets. Do not confuse a storyboard sketch with a finished branded template.
typography goal
Supply a single fictional hook and CTA with spelling locked. For example, use a made-up product label and a neutral CTA such as See the design brief rather than a real price, health claim, or legal promise. Specify maximum lines, contrast expectation, and safe-margin rule. Require reviewers to transcribe all generated text exactly. Image generators can produce persuasive near-matches; one extra or missing character is a failure if the copy is regulated or client-approved.
source-image permission.
Choose two images with permission for generative transformation and sharing. A fictional APOB influencer or an original APOB portrait can be a source option if its rights fit the use. Record provenance. No source image is included in this draft.
Turn one sketch into two different briefs.
Hold the sketch, hierarchy, exact hook/CTA text, target canvas, and evaluation rubric constant. Change only the licensed subject and its allowed product context. That lets reviewers ask whether the design survives a subject swap rather than comparing two unrelated prompts. Any account limitation, refusal, or failed generation belongs in the log.
subject A
For the first run, submit the original sketch, subject A's permitted portrait, the locked copy, and a concise style direction. Save the submitted inputs, output, timestamp, and any visible generation setting. Do not crop or retouch the result before grading it. Mark not run until those files exist. The launch page's example images cannot stand in for this output.
subject B
Repeat with subject B while keeping the sketch and layout instruction unchanged. The product may differ, but any changed required text must be recorded as a controlled variable. Label the two briefs clearly. If the account refuses an input or changes the canvas behavior, report that as an outcome rather than forcing an apparently parallel comparison.
unchanged sketch.
Compare output A and B against the same original drawing. Score hook zone, portrait zone, CTA zone, text accuracy, and crop safety as five independent checks. Attach the sketch and both outputs to a contact sheet. The pass threshold is editorial: define it before seeing the images. If a generated version omits the exact CTA, it should not pass merely because the overall hierarchy resembles the sketch.
Pass a shared prompt to a second creator.
The real handoff test asks someone other than the original author to open a sanitized prompt and make a new version with their own licensed photo. A prompt that works only because its author remembers hidden instructions is not reusable. Observe the recipient's questions and edits; those are useful data about the brief, not defects to hide.
shared prompt link
Create the link through the actual ChatGPT sharing interface after reviewing what it exposes. Do not publish a private conversation, customer face, token, or source asset accidentally. Record whether the recipient sees the prompt alone, the generated image, or more context. Store a redacted test link in the private evidence folder and check it from a separate account or permission context. This draft has no live shared link and therefore makes no access claim.
recipient changes
Ask the recipient to replace the subject and product details while keeping the zone constraints and CTA copy rule. Capture any clarification they needed. If the prompt requires author supervision, the handoff is not yet successful.
review notes.
Use a short interview after the run: Which instruction was ambiguous? Was the original sketch visible? Did the output preserve the reading order? Attach the recipient's output only with consent. Separate prompt usability from image quality.
Archive a redacted repeatability bundle.
A repeatable result consists of inputs, versioned instructions, outputs, and a reviewer decision. Keep a private master bundle and a redacted editorial bundle. If Sketch is unavailable, record that limitation.
source permissions
For each subject, store rights holder, permission scope, allowed distribution, expiry, and deletion contact. Include the sketch author's permission and the product image license. The public article can describe the process without revealing the original files. If an APOB-generated image is used as a separate source, identify it as such; do not call it a ChatGPT result.
prompt version
Version the exact prompt, sketch, locked text, comments, and sharing state. Use an ID such as sketch-poster-v1 and log every revision. A later improvement is a new run, not evidence that the first run worked. Save both successful and failed outputs so the claimed result can be checked. Remove personal data before a prompt becomes public.
Add a small manifest to the redacted bundle: sketch version, input-license IDs, subject A and B output IDs, accepted prompt version, the second creator's questions, and an access-expiry review date. A missing output should be recorded as missing, not replaced with a launch screenshot. This makes the exercise repeatable without publishing private portraits or implying that a share link transfers every hidden instruction from the original chat.
APOB asset option.
APOB can be a separate route to create a licensed fictional portrait or image source when the brief needs one. The APOB AI Image Generator and ChatGPT Images are distinct products; there is no claimed native GPT Image 2.5 integration. Choose the tool based on the asset rights, controls, and actual test result. For now, the design-tool verdict is intentionally open: a reusable sketch workflow must be demonstrated by the two-subject run and the second-creator handoff, not by an announcement.
If the handoff succeeds, publish only the redacted sketch, prompt, and acceptance notes that the rights holders have cleared. If it fails, state which instruction broke and preserve the failed version for the next iteration. A creator should be able to look at the bundle and repeat the layout experiment without access to the original author's private chat. That is the practical difference between a shareable idea and a promotional example that works only once.
