Illustration Production
Purpose
Create controlled, reviewable illustration assets and image-generation prompts for BaseCamp.
Core principle:
Illustration generation is not a prompt. It is a controlled visual production pipeline.
Trigger Cues
Use this skill when the user asks for any of the following:
- "make an illustration", "generate an image", "create a visual", "make a cover", "make a poster", "make a comic panel"
- "write an image prompt", "improve this image prompt", "prompt for OpenAI Image 2.0", "prompt for DALL-E", "prompt for Nano Banana"
- "use our house style", "make it match our style", "style lane", "style guide", "taste", "art direction", "visual philosophy"
- "use these images as reference", "what examples do we have", "accepted/rejected examples", "why did this output fail"
- "patient education visual", "waiting room comic", "AI Beyond Chatbot illustration", "presentation visual", "AED cover"
Do not use this skill for ordinary prose editing, non-visual research, deterministic UI implementation without custom illustration needs, or simple file organization unless visual taste/style/prompting is part of the task.
Canonical Guide
Read the guide first for any new or materially changed visual task:
references/illustration-guide.md- For a quick visual overview of what each style lane represents, read
references/style-lane-gallery.md.
Use existing project-local rules when they are more specific:
- project
AGENTS.md,STATUS.md, and local prompt files - established cover, comic, deck, or illustration design docs
- AI Beyond Chatbot examples:
references/ai-beyond-chatbot/illustration-prompts.md
Intake Gate
Ask a specific intake before creating or materially revising an illustration, prompt system, style lane, or visual asset.
Ask 4-5 focused questions, usually:
- Purpose and audience: what should this image do, and who is it for?
- Asset type and destination: cover, scene card, slide visual, poster, infographic, website asset, etc.; where will it live?
- Style lane or reference: use an existing lane/example, or create a new direction?
- Constraints and guardrails: text policy, medical/ethical concerns, character consistency, privacy, brand limits, size/aspect ratio.
- Autonomy level: should the agent choose and execute, or return options for sign-off first?
Do not implement skill file changes without intake. For ordinary asset work, if the user explicitly says to proceed with defaults, use the guide defaults and continue.
Style Lane Registry
Current lanes:
Warm Editorial Analog Clarity: default house style for editorial/conceptual visuals.Patient Education Indie Comic: clinic/patient education scenes, especially Waiting Room Comics.Minimal Academic Cover: AED-style abstract specialist cover art.Swiss Editorial Clinical Flat: clinical-academic slide illustrations and workshop visuals.Semantic Typography Concept Poster: typography-led concept posters where a word or phrase becomes the image structure.- Option:
Geometric Typographic Concept Poster, for modern flat geometric posters where the core word becomes a semantic visual symbol through type, grid, color fields, transparent layers, and minimal symbolic geometry. - Template:
references/semantic-typography-geometric-poster-template.md - Example:
references/semantic-typography-geometric-poster-shenjingneike.md - Example:
references/semantic-typography-epilepsy-poster-example.md - Example:
references/semantic-typography-neurology-poster-example.md
- Option:
Do not invent a new lane casually. Promote a new lane only when it has a repeated use case, accepted example, descriptor, avoid list, and a reason existing lanes do not fit.
Workflow
- Read the smallest relevant context.
- For new work, read the canonical guide.
- For project work, read the nearest
AGENTS.mdand projectSTATUS.md. - For a named lane, read that lane's canonical source.
- Classify the task.
- new asset
- prompt only
- style direction
- iteration on existing asset
- review/rejection pass
- lane creation
- Select one style lane.
- Pick one lane by default.
- Do not blend lanes unless the user explicitly asks or the task clearly requires it.
- State the chosen lane and why it fits.
- Choose the tool path.
- Default to OpenAI Image 2.0 for visual asset production through the Codex-native image generation tool. This includes illustrations, covers, posters, visual examples, style-lane examples, and typography-led concept art.
- When the user asks for Image 2.0, call the native Image 2.0 generation path directly. Do not require or check
OPENAI_API_KEY, and do not route through CLI/API wrappers or substitute non-native generation paths. - If the native Image 2.0 tool is not exposed or callable in the current session, stop and report that exact tool-surface blocker. Do not substitute another renderer or create a lower-quality fallback unless the user explicitly authorizes that fallback after being told it is not Image 2.0.
- Use Gemini / Nano Banana only when requested, when a verified project path requires it, or when OpenAI Image 2.0 is unsuitable; Gemini should run through the project-approved script/API path with credentials from environment, local untracked config, or macOS Keychain.
- Do not default to deterministic HTML/SVG/PNG for illustration or poster art just because text or geometry is involved. Use deterministic code/SVG/HTML/CSS/slides only when the user explicitly asks for a deterministic/editable pathway, or when exact labels, charts, diagrams, UI, or production typography are the deliverable.
- Label references by role.
- style reference
- character reference
- composition reference
- palette reference
- mood reference
- medical/device reference
- negative reference
- Draft the production prompt.
- Include purpose, audience, asset type, style lane, composition, palette, text policy, constraints, and avoid list.
- Use the guide's prompt architecture.
- Keep important text outside generated artwork unless explicitly approved.
- Run the pre-generation gate.
- purpose clear
- one lane selected
- one visual idea
- text policy explicit
- reference roles labeled
- medical/ethical guardrails explicit when relevant
- output unit appropriate for the model
- Generate, draw, or hand off.
- For raster image generation, use the Codex-native OpenAI Image 2.0 generation path and save accepted outputs into the project workspace.
- Never answer an Image 2.0 request with a non-native substitute.
- If native Image 2.0 generation is unavailable, hand off with the final prompt set, reference roles, and blocker instead of silently downgrading.
- For prompt-only tasks, deliver the prompt plus review checklist.
- For deterministic visuals explicitly requested by the user, build the visual in code or the appropriate document/presentation tool.
- Review the output.
- Check purpose fit, lane consistency, composition, palette, character consistency, bias/authority drift, text policy, medical plausibility, and thumbnail/distance readability.
- Save accepted project-bound assets in the workspace or project folder, not only in a temporary/generated location.
- Log rejected examples with a one-line failure mode when useful.
Iteration Path
For an existing asset, do not restart from scratch by default.
- Identify the accepted lane and source prompt.
- Name the exact failure: composition, palette, character drift, text, medical cue, density, or tone.
- Make one targeted change.
- Re-run the review gate.
Output Contract
When producing prompts or assets, report:
- selected style lane
- model/tool path
- reference roles used
- final prompt or prompt delta
- review checklist result
- saved asset paths, if any
Keep the response concise unless the user asks for a full art-direction packet.