Warm Humanist Editorial Art

Create an original editorial illustration with a clear idea, a visible action, and disciplined emptiness. The defining system is not “rough doodles.” It is a precise relationship between two graphic languages:

• near-black embodied gesture: hand, face profile, frame, axis, ribbon, tool, or boundary;
• warm off-white abstract system: nodes, branch, graph, path, molecule, book, aperture, or clustered form;
• one flat background field: muted and normally texture-free.

The image succeeds when the black gesture and off-white system make contact in a way that communicates the article thesis within two seconds.

Use the user's language for discussion. Write generation prompts in the language that performs best for the active image tool; default to precise English.

Do not use this skill for ordinary diagrams, realistic portraits, generic cartoons, logo design, or unrelated image editing.

Portable execution

Treat the directory containing this SKILL.md as the skill root. Resolve every bundled file relative to that directory, regardless of the current working directory or client. Before running a bundled script, change the process working directory to the skill root or pass the script's absolute path.

Use an available Python 3 interpreter (python3, python, or the client-provided Python runtime). The scripts use only the Python standard library. If Python is unavailable, follow the documented workflow manually and say that the deterministic compiler or storyboard check was not run.

Non-negotiable invariants

In signal-gesture mode, preserve all of these unless the brief materially requires an exception:

1. One specific editorial thesis, not a broad topic.
2. One concrete visual verb: trace, support, tune, open, hold, mediate, compare, release, or coordinate.
3. One dominant contact point between the black and off-white systems.
4. Two to four semantic components.
5. A flat background with no gradient, shadow, glow, lighting model, or 3D depth.
6. Near-black agency and warm off-white abstraction remain visually distinct.
7. Primary black path width is normally 1.6–2.4% of the short edge.
8. Foreground pixel occupancy is normally 9–18%, while the distributed silhouette may span 55–90% of a frame dimension.
9. No unrequested text, logo, watermark, stock icon, or copied arrangement.
10. The contact point remains readable at 128 px and survives every required crop.

Do not add paper grain by habit. Minimal-mode texture defaults to zero. Material texture is optional only when it carries meaning.

Determine the deliverable

Infer the required output before designing:

• finished image: generate, inspect, repair, and present the best result;
• prompt package: return art direction, generation prompt, exclusions, output spec, and two repair prompts;
• visual system: define concept families, ratio behavior, palette tokens, invariants, and consistency tests;
• reference analysis: describe transferable properties, then redesign the metaphor, pose, topology, crop, and silhouette;
• revision: identify the highest-impact failure and change only that variable first.

Ask one clarification only when a missing fact would change the concept itself. Otherwise state a compact assumption and proceed.

Workflow

1. Parse the brief

Extract or infer:

• subject and actual claim;
• actor or force;
• affected system or object;
• visual verb;
• tension, limit, consequence, or asymmetry;
• intended placement and aspect ratio;
• text-safe area;
• mandatory and forbidden content;
• palette constraints;
• whether hands, faces, archival material, or labels are allowed;
• whether the result is one image, a responsive family, or a reusable system.

Use CROP_AND_FORMAT.md when placement or multiple formats matter.

2. Write the editorial thesis

Privately reduce the brief to:

> [actor or force] performs [visual verb] on [system or object], while [tension or consequence] remains visible.

Weak: “AI safety.”

Strong: “A human-held boundary lets one useful path pass while keeping the wider system constrained.”

Reject theses that are merely nouns, slogans, or generic optimism. Do not expose private chain-of-thought; share only the concise final art direction when useful.

Read SEMANTIC_ENGINE.md for verb selection, metaphor distance, and cliché replacement.

3. Build three concept candidates

Create three candidates with genuinely different mechanics, not cosmetic variants. For each candidate define:

• black embodied element;
• off-white system;
• exact contact geometry;
• composition pattern;
• why the interaction maps to the thesis;
• main ambiguity or cliché risk.

Score each candidate from 1–10 on thesis clarity, specificity, silhouette, crop resilience, originality, and model reliability. Select the strongest total, but reject any candidate with thesis clarity below 8 or originality below 7.

Prefer the simplest candidate that preserves the tension. More objects do not make the idea smarter.

Use COMPOSITION_LIBRARY.md and, for hand-led concepts, HAND_GRAMMAR.md.

4. Lock the art-direction spec

Define before generating:

thesis: "one sentence"
visual_verb: trace
mode: signal-gesture
pattern: SG-04
black_element: "a cropped symbolic pinch gesture entering from the right"
offwhite_system: "a sparse branching path with five unequal nodes"
contact_point: "thumb hook and index loop overlap one connector"
ratio: "16:9"
background: "dusty-blue #C3D6D3"
ink: "near-black #141413"
offwhite: "warm off-white #FAF9F5"
primary_stroke_percent_short_edge: 1.8
occupied_foreground_percent: 13
semantic_components: 3
text_policy: none
forbidden: [robot, circuit-board brain, gradients, shadows, copied composition]

Use templates/art-direction.yaml or compile a validated package from the skill root with:

python3 scripts/build_prompt.py \
  --subject "model interpretability" \
  --thesis "a researcher traces one inspectable route through a hidden reasoning system" \
  --verb trace --pattern SG-04 --ratio 16:9 --palette dusty-blue

5. Compile the prompt for the active tool

Prompt in this order:

1. thesis and visual action;
2. black element, off-white system, and exact contact;
3. composition pattern, occupancy, negative space, and crop-safe placement;
4. line and shape behavior;
5. flat palette and text policy;
6. hard exclusions;
7. interface-specific instructions.

Never use a company name as the operative style instruction. If the user says “Anthropic-like,” translate it into the visible rules above, then remove the brand phrase from the production prompt.

Read MODEL_ADAPTERS.md and PROMPT_PATTERNS.md.

6. Generate the first draft

When an image tool is available:

• generate at the final aspect ratio rather than relying on a destructive crop;
• begin with one strong draft or a small set of conceptually distinct drafts;
• keep the same thesis and contact point across variants;
• do not request decorative detail to “make it interesting”;
• do not claim success until the image itself passes the gate.

When no image tool is available, return a complete prompt package and say plainly that no image was generated.

7. Inspect with four tests

Inspect in this order:

1. Two-second test: is the action understood without reading the title?
2. Contact test: can one point be named where the black action changes the off-white system?
3. Thumbnail test: at 128 px, do the action and silhouette remain readable?
4. Crop test: does every required format preserve the contact and at least 70% of the meaning-bearing silhouette?

Then inspect anatomy, accidental text, object count, line width, off-white outlines, flatness, clichés, and reference distance.

8. Repair one failure at a time

Use the smallest viable edit:

• unclear thesis → replace the metaphor, not the texture;
• weak contact → move or overlap the two systems explicitly;
• malformed hand → redraw only the black gesture using the symbolic loop grammar;
• crowded frame → remove secondary marks before shrinking the main subject;
• polished-vector look → introduce controlled low-frequency irregularity, not noisy sketch marks;
• outlined off-white system → remove its black perimeter while preserving the black gesture;
• crop failure → reposition the contact first, then reflow branches;
• generic AI imagery → replace the noun with a domain-specific action.

After two failed hand repairs, switch to a hand-free black force such as a ribbon, aperture, axis, frame, or profile. Do not burn unlimited rounds on anatomy.

Read ITERATION_PLAYBOOK.md.

9. Apply the final gate

Score with QUALITY_RUBRIC.md.

Finalize only when:

• weighted score is at least 9.2/10;
• no hard-failure cap applies;
• the dominant contact is clear;
• the final crop is safe;
• the image is materially original;
• no obvious repair remains.

A 10/10 image has no visible defect and no shape that fails to support the thesis. “Looks similar to a brand” is not a quality criterion.

Modes

Signal gesture — default

Sparse symbolic cover using near-black agency, warm off-white abstraction, and one flat field. Use for most research, product, science, safety, policy, and conceptual article images.

Archival evidence

Redrawn public-domain-looking source language plus a new encoding system. Use only when history, evidence, taxonomy, or measurement is central. Texture may be meaningful here, but must not become nostalgia decoration.

Explanatory note

Two to four sequential states or restrained panels. Use when the reader must understand a process rather than absorb one symbolic claim.

Originality and reference handling

Read REFERENCE_POLICY.md whenever a user supplies a reference or requests an exact match.

Preserve only broad properties such as line weight, palette restraint, spatial density, and humanist conceptual tone. Change the metaphor, object combination, pose, contact geometry, topology, crop, and silhouette. Never copy a logo, wordmark, proprietary character, or distinctive composition.

Deliverables

When generating an image, provide the image and one sentence naming the metaphor. Add no long explanation unless asked.

When returning prompts only, use:

ART DIRECTION
[one compact paragraph]

GENERATION PROMPT
[one production-ready prompt]

EXCLUSIONS
[negative prompt or hard exclusions]

OUTPUT SPEC
[ratio, dimensions, crop-safe zone, palette, text policy]

REPAIR PROMPTS
1. [most likely gesture correction]
2. [most likely simplification or crop correction]

For a reusable system, also include 3–6 concept families, ratio rules, palette tokens, and a consistency checklist.

Supporting files

Load only what the task requires:

• STYLE_FORENSICS.md: observed visual grammar and corrected misconceptions.
• STYLE_SPEC.md: numeric line, shape, palette, density, and texture rules.
• SEMANTIC_ENGINE.md: thesis-to-metaphor method and cliché controls.
• HAND_GRAMMAR.md: reliable symbolic hand construction.
• COMPOSITION_LIBRARY.md: ratio-aware composition skeletons.
• CROP_AND_FORMAT.md: responsive crops and export planning.
• MODEL_ADAPTERS.md: prompt behavior by image-tool interface.
• PROMPT_PATTERNS.md: base and repair templates.
• ITERATION_PLAYBOOK.md: visual diagnosis and targeted fixes.
• QUALITY_RUBRIC.md: weighted 10-point evaluation.
• REFERENCE_POLICY.md: originality and reference-image rules.
• EXAMPLES.md: worked briefs and compiled directions.
• CALIBRATION_NOTES.md: source observations and measurement limits.
• VALIDATION.md: package-level validation record.