---
name: agy-create-verified
description: "Operational guide for iterative image generation with visual and deterministic programmatic verification and edits."
---

# Verified image workflow

Use `agy_create_verified` only when a generated image must be checked and possibly corrected across iterations. Use `agy_generate` for one-shot or purely aesthetic work.

## Prepare the request

Separate requirements into:

- **Semantic:** subject, pose, style, mood, scene.
- **Layout:** aspect ratio, placement, count, alignment, spacing.
- **Measurable:** dimensions, dominant colors, geometry, contrast, text regions.

Write acceptance criteria as explicit pass/fail statements. OpenCV can verify measurable criteria; Gemini judges semantic criteria. Do not claim a subjective requirement is objectively verified.

## Deterministic edits

Use `editInstructions` for a repeatable transformation such as resize, crop, color correction, thresholding, masking, watermark placement, or geometric cleanup. Use `editScript` only for an explicit trusted script.

The edit script contract is:

- read input from `sys.argv[1]`;
- write only the output image to `sys.argv[2]`;
- use only `cv2`, `numpy`, `PIL`, `json`, `sys`, and `math`;
- do not use network, subprocesses, dynamic imports, or arbitrary filesystem access.

The programmatic edit runs before visual/OpenCV verification, so its output—not the raw generated image—is the candidate for the next iteration.

## Budget and interpretation

- Use `maxIterations=2` for simple corrections, 3 normally, and 4–5 only when justified.
- Keep `useOpenCV=true` when at least one criterion is measurable; otherwise it adds cost without reliable evidence.
- A `PASS` is valid only for the final candidate and available checks.
- A `FAIL` result is the latest candidate, not a compliant image. Report unmet criteria explicitly.
- Inspect the per-iteration report for edit errors, unchanged progress, regressions, and failed OpenCV output before presenting the result as verified.