Thalarch Visual QA

The final pixels are the source of truth for visual claims.

1. Derive a visual checklist

Convert the user's visual requirements into checks such as:

• required subject/content;
• composition/framing;
• exact text;
• style/brand match;
• palette;
• dimensions/aspect ratio;
• alpha/transparency;
• crop/safe zones;
• elements that must be absent;
• invariants preserved from the baseline;
• responsive/runtime states.

Use PASS, FAIL, UNVERIFIED.

2. View the whole artifact first

Before zooming into details, inspect:

• hierarchy;
• balance;
• focal point;
• crop;
• readability;
• contrast;
• overall visual coherence;
• whether it obviously violates the requested direction.

For polished/professional work, also ask whether there is any obvious visual weakness that would benefit from one small targeted improvement: weak focal hierarchy, awkward spacing/crop, unnecessary clutter, or a brand/reference mismatch.

Do not turn this into a style blacklist. Gradients, glow, symmetry, bokeh, dramatic lighting, 3D, minimalism, or decorative effects are valid when they work for the image.

Then inspect details.

3. Mechanical image probes

Use the bundled read-only scripts when relevant:

python scripts/image_probe.py <image>
python scripts/image_compare.py <baseline> <candidate> --out <diff.png>

They can prove properties such as dimensions, format, alpha, and same-size pixel change statistics. They cannot decide whether the design is beautiful.

4. Exact text

If exact visible copy matters:

• read the rendered text from the image itself;
• compare character-for-character with the required copy;
• inspect line breaks when layout matters;
• flag extra or hallucinated text.

Do not infer text accuracy from the generation prompt.

5. Image-edit preservation

For "change only X" edits:

1. compare against the original;
2. inspect requested region;
3. inspect several supposedly unchanged regions;
4. use a visual diff when dimensions align;
5. distinguish intended global effects from collateral drift.

Check for unintended changes to:

• identity;
• pose/geometry;
• framing;
• background;
• lighting;
• color grading;
• text/logo;
• texture/sharpness.

6. Artifact quality

Inspect for common generative/editing failures when relevant:

• malformed anatomy;
• duplicated objects;
• warped geometry;
• broken perspective;
• fake/garbled typography;
• halos and bad masks;
• inconsistent lighting/shadows;
• seams from compositing;
• unintended watermarks/logos;
• compression/upscale artifacts;
• banding;
• transparent-edge contamination.

Do not mechanically hunt for every category if it does not apply.

7. Brand review

For brand assets compare against the actual brand contract:

• mark integrity;
• color roles;
• typography character;
• spacing/safe area;
• shape language;
• imagery treatment;
• recognizability at small size.

Novelty is not automatically brand consistency.

8. Web/UI visual QA

For implemented UI, pair this skill with real browser/device evidence.

Compare screenshots against:

• design-system contract;
• reference/mockup if present;
• compact and desktop layouts;
• hover/focus/open states where visually meaningful;
• long text/empty/error states when relevant.

Check:

• hierarchy;
• rhythm;
• clipping;
• alignment;
• responsive reflow;
• image crops;
• typography;
• contrast;
• obvious accessibility regressions.

A generated mockup is not evidence of the implemented UI.

9. Annotated findings

When a visual defect is hard to describe precisely, create an annotated copy of the screenshot/image with numbered callouts. Keep the original untouched.

Use the annotation only as evidence; fixes must target the actual source asset or implementation.

10. Convergence rule

When a candidate fails:

• identify the smallest visual delta required;
• preserve all already-passing constraints;
• request one targeted edit pass;
• re-run only checks invalidated by that change plus a whole-image sanity check.

Do not restart the creative direction for a local defect. If the image is already strong and meets the contract, a clean PASS is preferred over needless regeneration.

Output

Return:

Requirement | PASS / FAIL / UNVERIFIED | Evidence

Then separate:

• blocking defects;
• optional polish;
• exact next edit, if another pass is necessary.

A clean pass is valid. Do not manufacture criticism.