Skip to content
← All articles

AI & automation

Version your AI ad prompts so the next test is reproducible

Keep prompts, reference images, model settings and approved exports together. A practical HeliosGen and Kie.ai workflow with a downloadable version log.

By WhatWins Team · AI-assisted

In this article

The most attractive output in a folder is not necessarily the most useful one. If nobody can find its prompt, product reference or model settings, a small correction can become another round of guessing.

Use the prompt and asset version log alongside the HeliosGen and Kie.ai setup guide. This is a record-keeping workflow you can apply to an existing tool. It does not require a new automation service or a paid generation run to prepare.

Give the concept and the export different identifiers

A concept describes the advertising idea. An export identifies a particular file. Keep both. Otherwise a new crop or corrected label can accidentally become a new creative concept in the reporting sheet.

For a fictional backpack campaign, use a concept identifier such as commute-access and an asset identifier such as commute-access-v03. The third version might only correct the background. That does not mean the customer objection or message changed.

Write the naming rule in the shared folder and use it consistently. Avoid customer information, API keys and private account details in filenames. The objective is to let a colleague identify the asset without opening every image, not to encode the whole campaign in one string.

Capture the inputs before running the model

Kie's Nano Banana 2 documentation describes request inputs including prompts, image inputs, aspect ratio, resolution and output format. Record the settings actually used by your provider and interface, rather than assuming that the model name alone describes the request.

The version log should contain the concept, prompt version, reference filename, model, output settings and eventual task identifier. Keep the original prompt text in a neighboring plain-text file. If a tool does not expose a seed or another control, record it as unavailable; do not invent a value to make the record look complete.

Preserving these inputs makes the procedure traceable. It does not guarantee pixel-identical regeneration: providers, model behavior and unsupported controls can change. Keep the actual approved export as the source of truth for what was reviewed.

Change one useful instruction at a time

Here is an original drafting exercise. No outputs or campaign results were generated for it.

Prompt v01: Place the supplied backpack reference on a neutral desk beside a closed laptop. Preserve the visible product shape, color and strap construction. Do not add text, logos or accessories to the product.

Prompt v02: Keep the v01 instructions. Move the laptop farther from the bag so it does not hide the lower pocket. Leave clear space above the product for copy added later in the design file.

The reason for v02 is specific: the first composition could obscure an important product feature. A reviewer can check whether the change addresses that issue. “Make it more premium” would be harder to evaluate because it does not define a visible correction.

When a version changes the setting, product angle and message at once, record those changes honestly. It may be a useful new concept, but it is not a clean comparison of one instruction.

Keep generation, editing and approval separate

An image returned by a provider is a generated asset. A crop or retouch creates an edited asset. Approval is a decision about a specific version and use. Give each stage a place in the log so “generated successfully” cannot be mistaken for “approved for an ad.”

StageRecordExit condition
GenerationTask identifier and original outputA file exists and can be reviewed
EditingParent asset and edited exportRequested corrections are complete
ApprovalReviewer, date and intended placementProduct, claims and crop checks pass

Run the AI image quality checklist on the edited export. If you change a product detail after approval, return that asset to review. Avoid overwriting the approved filename with a different image.

Package a small handoff that another person can use

A practical handoff contains the final export, the prompt file, the reference filenames, the version log and the intended advertising question. Include a short note describing any unresolved issue. Exclude credentials and private provider links that the recipient cannot safely use.

For the HeliosGen starter, keep a copy of the canvas before changing branches. Follow the existing guide's version-specific import instructions, and inspect shared files before passing them to another person. A saved workflow and a finished creative answer different needs.

When taking approved assets into supported Meta or Google Ads work in AdRails, keep the same asset identifiers in the brief. That makes the next review easier to connect to the exact creative that was approved.

Provider documentation checked October 15, 2026. The naming convention, prompt example and handoff procedure are WhatWins recommendations. No paid generation or ad-account action was performed for this article.

Creative test matrix

A spreadsheet-ready worksheet for five creative directions, their budget, results and the next decision.

Get the free worksheet