Grant figure assets

Grant Figure Assets

Workflow

Use this skill together with imagegen for proposal figure materials.

1. Decide the asset strategy:
   • For layout exploration, generate one full reference infographic first.
   • For reusable HTML assets, prefer one icon per generation rather than cropping a gridded sheet when borders, labels, or uneven placement would contaminate crops.
   • Use a gridded sheet only for rough style exploration or when the user explicitly wants a sheet.
2. Generate raw images with imagegen.
   • Save raw outputs under a project-local directory such as figs/icon_assets/<batch>_raw/.
   • Do not leave project assets only under $CODEX_HOME/generated_images.
3. Tight-crop raw images with scripts/tight_crop_assets.py.
   • Use a white-background threshold to find the non-white bounding box.
   • Save cropped assets under a sibling directory such as figs/icon_assets/<batch>_tight/.
   • Create a contact sheet for visual QA.
4. Inspect the contact sheet with view_image.
   • Regenerate any asset whose concept is wrong, whose edges are cropped, or whose image has unwanted labels/borders.
   • Re-run the crop script only for the changed assets or the whole raw folder.
5. Report the saved paths and any assets that were intentionally superseded.

Prompt Rules

For JSPS/DC1-style monochrome assets:

• Use black line art on pure white background.
• Prefer crisp, print-safe, simple but recognizable illustrations over generic icon symbols.
• Ask for square canvas and no outer frame, no grid, no labels unless labels are intentionally part of the asset.
• State that the object should fill roughly 70-80% of the canvas with about 10% padding.
• For HTML reconstruction, keep text labels in HTML/CSS whenever possible and keep PNG assets label-free.
• If a generated sheet includes grid lines, labels, or inconsistent placement, switch to individual generation.

Single-object rule (mandatory for HTML reconstruction)

When the target figure will be reconstructed in HTML/CSS/SVG (as opposed to being used as a standalone illustration), every PNG asset MUST be a single, self-contained object. Do not pack a process or a flow into one PNG. This is the most common failure mode for proposal figures and is non-negotiable.

Concretely:

• One asset = one object. If the concept involves "input → process → output", it is three objects. Generate three separate PNGs (input object, process object, output object) and let HTML add the arrows between them.
• No arrows inside the PNG. Arrows are layout, not content. They belong in HTML/CSS/SVG, not in raster assets.
• No paired "before/after" depictions. A "compression" icon should not show a big document next to a small document with an arrow. It should be either the big document alone or the small document alone — pick one and generate the other separately if needed.
• No composed scenes. "Brain reading a document" = brain + document = two objects = two PNGs.
• No decorative satellites. An LLM icon should not be a brain with orbiting atoms / sparkles / dotted halos. It should be just the brain (or just the network graph) and nothing else.
• Style must match the existing batch. If the user references an existing icon set (e.g. 01_pos_system.png ... 12_model_size_package.png), every new asset must read at the same visual weight: same stroke width, same level of detail, same "single recognizable noun" framing. Look at one or two of the existing icons via view_image before drafting prompts.

If the user's asset list violates this rule (e.g. asks for "X → Y → Z" in one icon), push back and propose a decomposed list before generating. Do not silently generate a flow diagram and hope for the best — the resulting assets will be unusable for HTML reconstruction and will need to be regenerated.

Reference figure scale check

If a reference figure is provided (e.g. a GPT Image 2 mockup of the target infographic), measure how much pixel area each pictogram occupies in the reference. Pictograms that occupy ~80×80 px in the reference must be generated as simple, sparse symbols — not as fully-rendered illustrations scaled to 1024×1024 — because the generator will fill extra canvas with decorative elements (arrows, sub-parts, orbital decorations) that break the single-object rule. When in doubt, ask the generator for "sparse line art, 30%+ whitespace, single centered symbol, no secondary elements."

Read references/prompts.md only when you need reusable prompt templates for full figures, icon sheets, or individual icon regeneration.

Cropping

Run the script with uv so Pillow does not need to be globally installed:

uv run --with pillow python /Users/yosuke/.codex/skills/grant-figure-assets/scripts/tight_crop_assets.py \
  --input path/to/raw-assets \
  --out-dir path/to/tight-assets \
  --threshold 218 \
  --contact-sheet

Useful options:

• --padding 0: true bounding-box crop with no extra whitespace.
• --padding 8: small margin for icons that visually need breathing room.
• --square: center each crop on a square white canvas after tight cropping.
• --max-size 1024: downscale long edges after cropping.
• --threshold: pixels darker than this are treated as ink.

Use --threshold 218 as a default for generated white-background line art. Increase it if faint gray lines should count as ink; decrease it if background haze is being included.

Quality Checks

Before finishing, verify:

• No grid lines, crop guides, or sheet borders are present in individual assets.
• Important strokes are not cut off at the top, bottom, left, or right.
• The background is white and the asset is grayscale or black-and-white.
• The icon concept is specific enough for the proposal text, for example POS register rather than a generic database when the user asked for POS.
• The contact sheet makes all assets recognizable at the target display size.