visual-assets
Skill2 starsChanged 10 days agoMITView on GitHub
Produce project-ready Document Studio visual assets from page art direction: connected explanatory scenes, editorial illustrations, pictogram families and targeted image edits. Use when meaningful asset production is needed; return actual files, provenance and inspected placement regions without taking over document layout.
In this repository9 other files it gives its agents.
- business-japanese-translationSkill
- business-korean-translationSkill
- company-profile-analyzerSkill
- document-architectureSkill
- document-designSkill
- document-researchSkill
- document-writingSkill
- html-to-pptxSkill
- report-html-pptxSkill
The file
Files
Meaningful document assets
The accepted guide baseline makes meaningful visual explanation routine in new documents. Follow concrete cover/body image jobs through to actual files and integration; adapt the scene to this topic rather than copying the guide paper metaphor. Preserve factual originals, native labels and concept status.
Consume the current brief and designer's concrete visual job. Do not restart document intake or invent an argument. Reuse content/visual-assets.json when present; its six core fields are asset_id, section_ids, purpose, semantic_constraints, production and delivery. Asset contract applies for integration/multi-asset jobs; a small standalone request can use the equivalent short brief.
Receive the supported meaning/subjects, final frame, image role, protected crop, view, quiet overlay region, adjacent caption, reference role, actual/concept state and selected expression. Resolve only a required original or a semantic gap that changes the visual. Routine aesthetic details are chosen and produced within the authorized request, without per-image approvals.
| Actual job | Production |
| Rich scene, original illustration or raster edit | Available native image-generation tool with its imagegen instructions |
| Exact simple diagram or geometric motif | Native SVG/HTML shapes; keep important labels/connectors in the document |
| Coherent pictogram family | Licensed supplied SVGs or a consistent vector extension |
| Actual logo, packaging, character or factual photo | Verified original with the accepted identity preserved |
| Exact quantitative chart | Design/data owner for editable accurate encoding |
If the selected job depends on convincing depth, architecture, perspective, organic subjects, material, lighting or an editorial scene, produce an actual image asset when generation is available. Do not downgrade it to CSS-like icon blocks merely because HTML is the final container. Conversely, do not generate an image for a text-first cover, exact chart or simple editable diagram. Mixed scenes can use a generated background and sourced identity assets with native text/motion overlays. See design's medium selection only when the method is ambiguous.
For v4.2, carry the designer's medium and reason from the existing page map into the six-field asset record only when production is needed. If image generation cannot finish, retain partial/blocked with the missing image job; never relabel unproduced concept art as a verified native diagram. A real logo or package remains blocked until an authorized original is available, even if a generated lookalike feels plausible. Exact labels/data and motion belong to HTML overlays, not the generated pixels.
Use expression from production.expression or brief.design_expression. For body jobs, the asset can connect subjects, support one paragraph, dominate a mosaic or supply a whole/detail view. Carry that relationship into production instead of using a generic cover prompt. Connected scenes need meaningful spatial grouping; an isolated stock-like object is insufficient for an integrated explanatory world. Quiet/inventive and rich/conventional are valid combinations.
Apply the chosen explanation job to the actual body asset: annotated subject/detail, mechanism, activity sequence, before/after concept or connected scene. Produce and place useful images instead of returning an image suggestion only. Repeated page decoration and a cover image cannot substitute for body explanation. Preserve factual originals, concept disclosure and native business text; coordinate exact diagram/chart jobs with design rather than baking them into an illustrative image.
Keep one compact family specification/reference anchor for projection, optical size, shape, palette, light and people proportions. Vary subjects and composition purposefully rather than appending every result as a style reference. Use style families, pictograms, image production or prompt patterns only for the current production issue. Dramatic light/texture/3D/color are available when useful; avoid making glossy floating objects, blank margins or pale cards universal defaults.
Native generation follows supported per-asset calls; do not invent seed/batch/style-lock parameters. Keep important titles, quantities, labels, conditions and meaningful arrows native in HTML. Explicit embedded text requires exact wording and actual inspection. Generated scenes never stand in for real facilities or completed activity. Detail crops from one image remain one source and cannot imply different sites/tests.
Inspect each actual returned image at intended size for subjects, relationships, palette/projection, transparency, focal/crop regions and unrequested additions. Inspect the integrated page when overlays affect meaning/readability. Report actual regions, not prompt-coordinate guesses. Repair only affected items using the stable anchor; preserve accepted original identity.
Copy usable outputs into the current project, normally html/assets. Return actual dimensions/format/transparency, source kind, prompts/reference roles, final paths and scoped review limits. Update delivery in the existing asset record and preserve unrelated fields. A preview/default generation directory alone is not a delivered asset. Reference provenance is conditional; no need to load all ESG sources or an offline comparison helper for ordinary production.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

