Thalarch Image

Images are first-class engineering artifacts. Treat them with the same discipline as code: define the contract, choose the right production path, preserve invariants, and verify the actual output.

1. Classify the visual task

Choose one primary mode before acting:

• inspect — understand an existing image or screenshot;
• generate — create a new raster visual from text/references;
• edit — change a specific part of an existing image;
• compose — combine multiple source images;
• vector — logo, icon, diagram, or geometry that should remain code-native;
• capture — browser/app/system screenshot used as evidence;
• compare — before/after or reference-vs-output fidelity;
• annotate — callouts, arrows, labels, highlights;
• optimize — format, size, alpha, crop, compression, or export cleanup.

Do not route every visual task through generative image creation.

2. Discover visual capabilities

Before choosing a production path, inspect what the current host/session actually exposes:

• image generation/editing;
• image viewing/inspection;
• browser screenshot/recording;
• device/emulator capture;
• filesystem/code-native SVG or canvas creation;
• deterministic image-processing scripts.

A capability available in Antigravity is not automatically available in Codex or Claude Code, and vice versa. Never fabricate a visual tool or subagent name.

If the acceptance criterion requires a capability that is unavailable, use the strongest safe fallback and keep the affected claim UNVERIFIED.

3. Choose the production path

Prefer the most deterministic medium that fits the deliverable:

• host-provided image generation/editing for raster generation and semantic image edits when an actual compatible tool is available;
• SVG / HTML / canvas / repository-native vector code when exact geometry, typography, editability, or scalable logos/icons matter;
• real browser screenshot tooling for web UI evidence rather than recreating UI as generated art;
• device/emulator capture for mobile runtime evidence when available;
• deterministic annotation/image-processing scripts for callouts, dimensions, alpha, crops, and measurable comparisons.

A generated bitmap is not a substitute for an SVG when the product needs a real vector asset.

4. Visual acceptance contract

For meaningful image work establish the fields that matter:

• purpose / destination;
• asset type;
• target file format;
• target dimensions or aspect ratio;
• safe zones / crop behavior;
• background and transparency requirements;
• exact text, if any;
• subject and composition;
• style / medium;
• palette / brand constraints;
• reference images and each reference's role;
• invariants that must not change;
• forbidden/unwanted elements;
• required variants;
• verification evidence.

If a field does not matter, leave it unspecified instead of inventing decorative requirements.

5. Reference-image roles

Never assume every attached image is an edit target.

Label references explicitly:

• edit target;
• style reference;
• composition reference;
• identity/character reference;
• brand/palette reference;
• content/source reference;
• comparison baseline.

When several images are used, preserve their numbering and role in every generation/edit brief.

6. Exactness hierarchy

Some properties need different proof:

• semantic appearance → visual inspection;
• exact wording → read rendered text back;
• pixel dimensions → metadata probe;
• transparency → alpha-channel check;
• before/after fidelity → same-size comparison and visual diff;
• brand palette → source/design tokens plus rendered inspection;
• layout geometry → prefer vector/code-native production or deterministic overlays when precision matters.

Do not claim a visual property from the prompt alone.

7. Asset hygiene

• Use semantic filenames.
• Preserve source/reference files unless replacement was requested.
• Do not overwrite the only original during experimentation.
• Keep exploratory outputs out of production asset paths.
• Do not silently change format, crop, color profile, or transparency.
• Avoid embedding secrets, local paths, private metadata, or unrelated EXIF data.
• For repository assets, inspect file size and intended delivery cost.

8. Handoff by capability

For generation/editing, use thalarch-imagegen plus a real host-supported image capability.

For final inspection, before/after comparison, dimensions, alpha, text, annotations, or visual evidence use thalarch-visual-qa and whatever actual image inspection/runtime tools exist.

For product UI combine with thalarch-ui, thalarch-browser-qa, and real browser/device evidence when those capabilities are available.

Do not phrase an unavailable handoff as if it executed. Missing capability becomes explicit UNVERIFIED evidence, not an invented specialist result.