Create Wildtrace Atlas

Treat this skill as a data-to-image compiler, not a style prompt.

Required resources

• Read references/style-language.md before art direction.
• Read references/specification.md before changing a board spec or accepting user data.
• Read references/composition-system.md when planning panel geometry.
• Read references/quality-system.md before approving or repairing generated images.
• Read references/model-adapters.md for the active image tool.
• Query the 120-method catalog with scripts/wildtrace.py catalog; do not load the entire JSON unless bulk catalog work requires it.
• Use assets/reference-board.png only as a visual-language reference. Never copy its layout or data.

Compiler workflow

Follow this sequence. Do not skip directly from a user request to an image prompt.

1. Resolve methods. Normalize chart names through the catalog.
2. Write data contracts. Use user data when supplied. Otherwise use catalog defaults and state that they are illustrative.
3. Plan the artifact. Produce a versioned JSON board spec.
4. Audit the spec. Fix deterministic errors before image generation.
5. Compile prompts. Generate one prompt per complete board from the audited spec.
6. Generate boards. Use one image-generation call per board. Never generate panels separately unless the user explicitly requests separate assets.
7. Audit pixels. Inspect the image with the quality rubric. Confirm method identity, data character, density, material language, and absence of unwanted UI or text.
8. Repair narrowly. Keep valid panels invariant. Change only the failed geometry or density, then inspect again.
9. Persist deliverables. Save every requested final board, its spec, compiled prompt, and audit report together.

Commands

Set SKILL_DIR to this skill folder, then use the standard-library CLI.

python "$SKILL_DIR/scripts/wildtrace.py" catalog --search sankey

Plan a board:

python "$SKILL_DIR/scripts/wildtrace.py" plan \
  --methods lollipop-chart,streamgraph,sankey-diagram \
  --seed 42 \
  --out work/board.json

Override catalog data:

python "$SKILL_DIR/scripts/wildtrace.py" plan \
  --methods lollipop-chart,sankey-diagram \
  --data work/data-overrides.json \
  --out work/board.json

Audit and compile:

python "$SKILL_DIR/scripts/wildtrace.py" audit --spec work/board.json --strict --out work/audit.json
python "$SKILL_DIR/scripts/wildtrace.py" compile --spec work/board.json --out work/prompts.txt

Or build the complete pre-generation evidence bundle in one command:

python "$SKILL_DIR/scripts/wildtrace.py" bundle \
  --methods lollipop-chart,streamgraph,sankey-diagram \
  --provider your-provider \
  --model your-model \
  --dir work/wildtrace-run

This writes the spec, deterministic audit, compiled prompts, method-to-board index, hashed run manifest, and a visual-audit skeleton. Fill the visual audit only after inspecting the generated pixels.

Plan all 120 methods as twenty six-panel boards:

python "$SKILL_DIR/scripts/wildtrace.py" batch --all --panels-per-board 6 --seed 42 --out work/all-120.json

Compile a targeted repair prompt:

python "$SKILL_DIR/scripts/wildtrace.py" repair \
  --spec work/board.json \
  --report work/audit.json \
  --out work/repair.txt

Run the built-in integrity check when the skill changes:

python "$SKILL_DIR/scripts/wildtrace.py" doctor

Non-negotiable invariants

• Encode data before choosing decorative geometry.
• Use exactly one visualization grammar per panel.
• Keep unsorted order unless ranking is intrinsic to the method.
• Preserve at least one meaningful irregularity per panel.
• Keep marks near 30–48% of the panel and preserve a calm field.
• Use two or more panel-background colors on multi-panel boards.
• Vary panel dimensions, anchors, and edges; do not use an equal card grid.
• Use color only for series, categories, direction, hierarchy, or emphasis.
• Omit titles, labels, numbers, legends, logos, watermarks, and UI chrome unless glyphs are intrinsic to the selected method.
• Reject icon-like symmetry, wallpaper density, glossy gradients, unrelated decoration, and stock-dashboard polish.

User data rules

• Never silently replace supplied values with prettier values.
• Never sort, normalize, aggregate, or smooth supplied data without stating the transformation and obtaining approval when it changes meaning.
• Preserve signs, order, missing values, units, hierarchy, source/target direction, and flow conservation when applicable.
• If the user supplies only method names, use catalog defaults and label the result as illustrative.
• If the user supplies more methods than fit one board, create a series of complete boards; keep a stable seed and shared material language.

Image-generation rules

• Treat user images as references unless the user explicitly requests an edit.
• State each input image's role in the prompt.
• For bitmap art, use the active image-generation tool rather than substituting SVG or UI code.
• Feed the complete compiled board prompt in one call.
• Do not add a brand name, logo, signature, or in-image title.
• Save the generated result outside temporary storage when it belongs to a project.

Pixel audit

After generation, inspect every panel and record pass/fail for:

1. method identity;
2. data-order and anomaly preservation;
3. one-grammar-only compliance;
4. sparse mark count and negative space;
5. unequal board geometry and multiple backgrounds;
6. tactile paper/ink material;
7. unwanted text, UI, decoration, logo, or watermark.

Use references/quality-system.md to score the result. Do not approve a board with a method-identity error. Repair that board only.

Deliverables

For a project-bound result, deliver:

• generated image files;
• board-spec.json containing actual methods and data;
• prompts.txt containing the compiled prompt set;
• audit.json containing deterministic and visual findings;
• run-manifest.json binding the run to a content hash, provider, model, and file set;
• visual-audit.json containing the completed 28-point pixel review;
• a concise method-to-board index for multi-board series.

The images are examples. The spec, catalog, compiler, and audit trail are the reproducible system.