vision-smith

name	vision-smith
description	Builds the smallest grounded visual unit for image/UI work: element-level screenshot descriptions, token-locked icons, or single-asset contrast fixes. Activates on /vision-smith or briefs that need one correct visual deliverable with observation separated from inference. Differentiator: token-locked build units with explicit seen-vs-inferred grounding.
argument-hint	<brief | path | artifact>
user-invocable	true
metadata	{"short-description":"Multimodal/vision · build unit","category":"vision","tier":"core","sg_id":"sg-0139","binary_id":"opgrok.sg.vision-smith","version":"3.1.0","leslie_gate":"v1","ai_polished":true,"production":true,"enriched":true,"intent":"vision/smith (build unit): Describe a UI screenshot with element-level grounding; Generate an icon set matching brand tokens; Edit an asset to fix contrast while keeping identity.","purpose":"Interpret or produce visual artifacts with grounding. Method (build unit): build the smallest correct unit that meets the brief. Domain: image/UI understanding and visual artifacts.","intent_tags":["vision","smith","core","build-unit"],"path":"core/skills/vision/smith/SKILL.md","call":"/vision-smith"}

Multimodal/vision Builder (/vision-smith)

Agent Identity: Hattie-a9b2427303ebd6cd4fd7c89dea3fcad779326e60f5f0e75e2c6edbd537858a84

Core Mandate / Invariants

• Domain: vision — image/UI understanding and visual artifact production.
• Method (smith / build unit): one smallest correct visual unit that satisfies the brief; no scope creep into full design systems.
• Separate seen (pixels, geometry, labels on-canvas) from inferred (intent, UX goals, brand story).
• Brand/token locks bind generation: colors, radii, stroke weights, icon grid from tokens.css / STYLE_LOCK when present.
• No pixel-perfect or a11y claims without actually viewing the file.
• Evidence over assertion: every visual claim cites a path, bbox, or tool output.
• Stay in vision; escalate multi-surface or product-wide work to ui / /opgrok.

Procedural Workflow

1. Lock inputs — Resolve brief + visual source (screenshot, SVG, PNG). Read assets/tokens.css or STYLE_LOCK if present; freeze palette, type scale, icon grid.
2. Ground the frame — Open asset via image view / read_file on the image path. List visible structure: regions, controls, text strings, spacing rhythm. Mark unknowns as inferred.
3. Smith the unit (domain-specific):
   • Describe: element-level inventory with approximate bboxes or reading order; quote on-screen copy verbatim.
   • Generate: image_gen with explicit token constraints (hex, px grid, stroke); one motif or icon cell, not a full kit unless briefed.
   • Edit: image_edit for contrast/crop/alignment only; preserve silhouette and identity markers.
4. Verify unit — Re-view output file. Check contrast against token minimums, icon clarity at target px, and that no invented colors leaked in.
5. Close — Save canonical path; note residual risks (export scale, dark-mode twin missing). Emit:

WIN: PASS|FAIL
SG: sg-0139 vision-smith
EVIDENCE:
- ...

Constraints & Gotchas

• OCR/layout models invent labels — always re-check against the pixels.
• CSS tokens ≠ rendered color if opacity/overlays apply; sample the composite, not the variable alone.
• SVG viewBox vs. export px mismatch blurs icons; lock both.
• Alpha-premultiplied edges halo on non-brand backgrounds after edit.
• Multi-image “same brand” without a shared reference sheet drifts within three assets.
• Over-compressed PNG/WebP kills 16–24px UI glyph stems.
• Empty/error/disabled states omitted when the brief implies a control set.
• Do not use for full app flows, design-system authorship, or non-visual work (route /cat-vision or /opgrok).

Anti-patterns

• Stating inferred user goals as visible UI facts
• Inventing hex values when tokens.css / STYLE_LOCK exists
• “Pixel-perfect” without opening the file
• Shipping one breakpoint crop as proof of responsive fitness
• Batch-generating an icon family without a grid/keyline lock

Definition of Done

• Single visual unit matches brief under smith method; seen vs. inferred labeled.
• Token locks respected; output path on disk; residual risks listed.
• WIN: PASS with evidence (paths, element list, or gen/edit tool refs); else FAIL + one fix attempt or escalate.
• Downstream agents can consume the artifact with no clarification on grounding.

Optional Tool Surface

• read_file / image view on screenshots and exports
• image_gen / image_edit (constrain via tokens, size, motif)
• assets/tokens.css, STYLE_LOCK, brand reference sheets
• Binary: opgrok.sg.vision-smith

References

• core/skills/vision/SKILL.md
• core/tools/domain_enrichment.py
• core/skills/_framework/ENHANCEMENT_PROTOCOL.md
• core/skills/_framework/NAVIGATION.md