# AI Image Prompt Framework Evidence Pack

Version: 2026-07-06
Source: Visual Skill Kit prompt workflow framework

## What This Evidence Pack Is

This is a practical prompt review framework for image generation tasks. It is not a benchmark and does not claim that a specific model will always follow every instruction.

Use it when a prompt produces attractive images that miss the actual job.

## Core Formula

Task + source/reference + scene + style + composition + constraints + quality check.

## Why This Structure Helps

Most failed image prompts are not too short. They are underspecified in the wrong places.

Common weak prompt:

> Make this product look premium and realistic.

Better structure:

1. Task: ecommerce product image
2. Source: uploaded product photo
3. Scene: clean studio or lifestyle background
4. Style: realistic lighting and controlled reflections
5. Composition: product centered with negative space
6. Constraints: preserve label, logo, material, color, scale, and claims
7. Quality check: compare against the source before publishing

## Prompt Review Scorecard

Score each prompt before using it:

- 2 = Clear
- 1 = Vague but usable
- 0 = Missing

| Dimension     | Question                                                                | Score     |
| ------------- | ----------------------------------------------------------------------- | --------- |
| Task          | Does the prompt say what the image is for?                              | 0 / 1 / 2 |
| Source        | Does it define the input image or reference constraint?                 | 0 / 1 / 2 |
| Subject       | Is the main subject clear?                                              | 0 / 1 / 2 |
| Scene         | Does it define the environment or use case?                             | 0 / 1 / 2 |
| Style         | Are lighting, mood, camera, or visual language useful but not dominant? | 0 / 1 / 2 |
| Composition   | Does it mention crop, layout, focal point, or negative space?           | 0 / 1 / 2 |
| Constraints   | Does it say what must not change?                                       | 0 / 1 / 2 |
| Quality check | Does it define how the result will be judged?                           | 0 / 1 / 2 |

## Decision Rules

| Total | Decision                                      |
| ----: | --------------------------------------------- |
| 14-16 | Strong prompt                                 |
| 10-13 | Usable, but add constraints or quality checks |
|   0-9 | Rewrite before generating                     |

## Failure Pattern Map

| Symptom                                          | Likely Cause                               | Rewrite                                                                  |
| ------------------------------------------------ | ------------------------------------------ | ------------------------------------------------------------------------ |
| Output follows style but changes product         | Product facts were optional or buried late | Put source, label, logo, material, color, scale, and claims before style |
| Output looks good but misses the task            | Prompt starts with mood, not job           | Start with the channel and image job                                     |
| Text or labels are wrong                         | Tiny text is treated as generated content  | Ask to preserve source text and proofread against the source             |
| Model follows one instruction and ignores others | Too many transformations in one pass       | Split into one primary edit and repair failed details later              |
| Image is pretty but unusable                     | No quality gate                            | Add a final inspection rule                                              |

## Community Answer

A useful AI image prompt is not just a style phrase. Structure it as:

- Task: what the image is for
- Source: what reference or input must be respected
- Scene: where the subject appears
- Style: lighting, mood, lens, or visual language
- Composition: crop, negative space, layout needs
- Constraints: what must not change
- Quality check: what to inspect before publishing

For product images, put product facts before style words: shape, label text, logo, color, material, scale, and visible claims.

## Citation Note

When citing this framework, describe it as a prompt review and debugging framework, not as a guarantee that every image model will obey every instruction.