Art

• Every image comes from the intent that matches the request, and its composition is reasoned from the specific argument or content it illustrates: scene concepts weighed against each other, every element in the frame carrying a narrative role, never a subject list dropped into a prompt.
• Every image holds its intent's technique and palette, or the style PREFERENCES.md sets, unless the principal names another style, and every person in it follows the figure treatment described in Gotchas.
• Every label, number, value and arrow in a diagram, stat card or annotated image is factually correct and spelled exactly as the source has it; one misspelled word, garbled digit or wrong hex value fails the image however good the art is. When the request asks for an image rendered from authored SVG, the delivered file comes from that SVG, never from an image model.
• An edit of an existing photo or reference image changes only what the request names: the subject's identity, and every element the request did not mention, read the same in the output as in the source.
• A set of images made for one deck or page reads as one set: the same medium, palette, ground and figure style as the style references the request names.
• Every image is rendered in a working directory outside Downloads (or the staging directory its intent names), passes the ArtDirection quality loop there, and reaches Downloads only through ArtDirection's Deliver.ts; nothing is copied into a project before that hand-off, except where an intent wires its own winner in without handing back.
• A finished file is verified, not assumed: it was looked at with the Read tool in this turn, file reports real RGBA where transparency was the point, its native dimensions fit the destination, and a blog header ships both the transparent inline image and the opaque -thumb.png and carries the signature stamp the Essay intent applies.
• A one-of-a-kind generation survives any background cut: its raw stays on disk until the cut PNG has been read and every subject in it confirmed present.

YouTube, channel and video thumbnails are not this skill: the video skill owns thumbnails end-to-end with its own vendored art intents and compositors (moved out of Art 2026-08-28).

Generate.ts intent gate

Generate.ts refuses to run unless it gets --intent=<name> (or the explicit --freeform-confirmed opt-out), and exits non-zero with the intent lookup table. The rule was markdown-only until it was silently ignored and one session produced a run of rejected diagrams (incident 2026-04-30); it now lives in code. The --intent=<name> flag asserts that the matching intent file was read and its prompt template followed: the intent templates encode the technique, palette, composition rules and validation gate the bare model fails to honor, and a freeform prompt shipped straight to Generate.ts is refused.

The one exception: the user explicitly says "freeform" / "skip the intent" / "just run Generate.ts directly with this prompt: ...". Then pass --freeform-confirmed, which logs the explicit opt-out to stderr for audit. Without that instruction from the user, pick the matching intent and pass --intent=<name>.

If no intent matches the request, stop and surface it to the user before generating, proposing (a) the closest existing intent, (b) Visualize.md as the generic catch-all, or (c) a new intent created first through the CreateSkill skill.

Output staging

Every generated image, in every intent, is rendered first into a working directory outside Downloads (for example scratchpad/art/<slug>/, written $ART_DIR in the commands below) or the staging directory its intent names, never directly into Downloads or a project's public/images/ directory: images pushed straight to repos have failed several times. Once the image has passed the quality loop below and the verification under Tool contract, it is handed off with bun ~/.claude/skills/ArtDirection/Tools/Deliver.ts --artifact <file> --to ~/Downloads. A file is copied into a project (e.g., cms/public/images/) only after that hand-off, and the WebP and thumbnail versions are made from the staged file. Whether the principal approves before that move is the intent's rule: the Essay intent wires its picked winner into the site without handing back (principal directive, 2026-09-22), and an intent that says nothing waits for the principal's review.

Quality loop

Nothing this skill produces reaches the principal as a first draft (principal directive, 2026-09-17). The artifact goes through Skill("ArtDirection"): a direction council before new or large work, then the cross-vendor critique panel (skills/ArtDirection/Tools/Critique.ts) scores it against the rubric and names defects, then this skill iterates on those defects until the bar or reports the shortfall. For still images, one round is one command: Tools/EvalRound.ts runs the asserts, blind describer and meaning ladder, cross-vendor critique, and flipped tournament in parallel, picks the winner in code, and takes 2 minutes or less (contract: Tools/EvalRound.help.md). Hand-over is skills/ArtDirection/Tools/Deliver.ts, which refuses without a passing critique. Record the principal's verdict with skills/ArtDirection/Tools/Calibrate.ts.

Vendors

Generate.ts ships only the models listed under Tool contract. Removed model names (gpt-image-1, gpt-image-2, the dual-provider compare mode) exit with an error pointing at nano-banana-pro. Which image vendors this install allows is the principal's call, recorded in <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Art/PREFERENCES.md § Vendors; read it before changing a model path. Do not add a vendor as a substitute when a round fails: adding a new vendor to any lane requires the principal's explicit approval.

bun ~/.claude/skills/Art/Tools/Generate.ts \
  --intent=<IntentName> \
  --model nano-banana-pro \
  --prompt "..." \
  --size 2K \
  --aspect-ratio 16:9 \
  --output "$ART_DIR"/<filename>.png

<IntentName> matches a file under Intents/ without .md. API keys are in <configRoot>/.env (~/.claude/.env by default; Generate.ts also checks ${LIFEOS_DIR}/.env first); Setup.md lists them.

The default model is nano-banana-pro unless PREFERENCES.md pins another; flux and nano-banana take an aspect ratio in --size, and the per-model sizes, the model-per-job table and reference-image use are in References/Models.md. A blog header adds --thumbnail; its command, transparency rules, verification and post wiring are in References/BlogHeaders.md.

User customizations live at <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Art/ (configRoot is $CLAUDE_CONFIG_DIR when set, else ~/.claude). When the directory exists, its files override the defaults here; when it does not, the skill defaults stand. First-time setup (accounts, keys, config, a proof it works) is in Setup.md.

• PREFERENCES.md - Aesthetic preferences (visual style and influences, line treatment and rendering approach, color palette and wash technique, character design, scene composition), default model, output location, and reference image locations with style examples by use case. Load the relevant user-provided references before generating.
• PREFERENCES.md § Typography - maps each type role the intents name (wedge-serif title and subtitle, geometric-sans labels, condensed-italic insights, hand-lettered display) to the install's own faces; when the section exists, prompts name those faces for those roles. § Color Palette, § Character Design, § Signature and § Vendors likewise replace the skill defaults.
• DiagramFonts.json - font files per type role, read by Tools/AnimatedDiagram.ts
• CharacterSpecs.md - Character design specifications
• SceneConstruction.md - Scene composition guidelines
• History.md - the specific work behind the dated gotchas

Without customization, the default aesthetic is production-quality concept art suited to editorial and technical content.

• Verify image dimensions match the target use before claiming done. Social cards (1200x630, usually built by the site from the post's largest image), blog headers, and listing thumbnails (512 wide) have different size requirements. A header that works on the blog may be too small to build a sharp OG/social card.
• When asked to use a specific image URL or file, use exactly that asset. Don't substitute similar images. Past rating-1 failures came from using wrong image assets.
• --remove-bg may produce black backgrounds instead of transparency. Verify transparent PNG output visually before deploying.
• RemoveBg.ts converts in place and zeroes RGB under removed alpha, so the original is unrecoverable. Running it on a .jpg replaces the file with the cut .png; if rembg eats an element (paper stacks, thin props, the DA-name signature), you cannot flatten the alpha back to recover the source and must regenerate (cost: a full new roll, 2026-08-16). cp the original before running RemoveBg.ts on a one-of-a-kind generation, every time.
• --thumbnail / --remove-bg can erase whole subjects in multi-subject compositions (2026-09-23, a two-subject blog header). rembg's default model keeps only the most salient object and fades the rest; on a two-ecosystem scene it kept one hub and fogged out the rival network and the person, and because the raw is deleted and RGB is zeroed under removed alpha the loss was unrecoverable. For any composition with more than one subject, keep the raw (copy it, or point REMBG_BIN at a wrapper that saves it first) and cut with rembg -m birefnet-general, which kept every element with a clean edge. Read the cut PNG and confirm every subject survived before judging it.
• --remove-bg is unsafe for thin-linework technical diagrams. rembg classifies thin black ink on a light field as "background" and strips it, leaving a near-empty ghost. Documented 2026-05-11 on a flowchart. Mitigations: (a) prompt for thick saturated linework first so rembg has a strong signal, or (b) skip --remove-bg entirely when the destination background matches the image's background (an opaque diagram drawn on the page's own ground, PREFERENCES.md § Color palette, composites with zero visible seam, no alpha needed).
• Logo fidelity breaks in 3D/perspective scenes even with a reference image. Documented 2026-06-11 on a brand wallpaper set: straight-on and macro scenes held the glyph topology, but the isometric 3D scene changed it (an open stroke closed, a separate element dropped). For any perspective/3D composition with a logo, add topology-locked negative language to the prompt that names each feature of the mark ("do not close the open stroke", "do not omit the separate element", name every stroke and terminal) on top of --reference-image, and vision-verify the topology specifically. The install mark and its topology rules live in <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Art/PREFERENCES.md § Wallpaper.
• nano-banana-pro "4K 16:9" is actually 5504×3072 (43:24, ~0.8% wider than 16:9), saved as .jpg even when --output says .png. Disclose the native ratio when the spec says 16:9, and probe the real filename before Read/delivery.
• Image models default every person to white, so every figure treatment is stated in the prompt (principal, 2026-07-12, 2026-09-20, 2026-09-23). <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Art/PREFERENCES.md § Character design sets the treatment and the sentence the prompts carry; without it the skill default is a varied cast whose skin tone, age and build each prompt names, spanning a natural range. Any multi-figure scene (essay headers, comics, visualizations) states the treatment explicitly, a group that defaults to one ethnicity is a hard failure, and the ArtDirection rubric scores a violation as a critical defect. See Essay.md HUMAN FIGURES block + Step 8 checklist.
• Every art request uses its matching intent and that intent's technique, plus <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Art/PREFERENCES.md when present, unless the principal names a different style (principal, 2026-09-24). Feedback about tone ("more positive", "less dark") is fixed inside the technique: lighter ground, warmer washes, smaller cool accent, a different scene. It is never a license to switch medium. A "more positive" redo once went out in a different painted medium under --intent=Essay and was rejected. The Essay contract's mandates and medium bans now refuse that prompt in Generate.ts.
• Essay/blog headers are signed with the DA name (getDAName() from hooks/lib/identity; principal, 2026-06-20 + 2026-07-09) in a small, integrated cursive signature hand. Generate.ts stamps it programmatically (Essay.md Step 7.1: the macOS cursive SignPainter-HouseScript face by default, ~3% of image width, semi-transparent charcoal, slight rotation, tucked into the composition's bottom-right); never prompt the signature into the model, because it garbles. A large print-letter signature reads as a label rather than a painter's mark; faces the principal has rejected for the signature are listed in <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/Art/PREFERENCES.md § Signature.
• Kept inline: the frontmatter eval contract, the four safety gates (intent gate, output staging, quality loop, vendors), the default Generate.ts command, and the cross-intent gotchas that RemoveBackground, TechnicalDiagrams, Comics, Visualize, Mermaid, Comparisons and the wallpaper intents point at as "SKILL.md Gotchas"; every run needs them.
• Image generation is Google plus Flux, never OpenAI — Nano Banana Pro (the default) and Nano Banana from Google, plus Flux; the last two run on Replicate. OpenAI image models are not used. Image skills and intents must not offer an OpenAI image option, and requests for one should error and point to Nano Banana Pro.

Example 1: Blog header image

User: "create a header for my AI agents post"
→ Essay intent
→ A charcoal-sketch header picked by EvalRound.ts, as a transparent PNG plus a page-ground -thumb.png staged outside Downloads, then optimized and wired into the post

Example 2: Technical diagram with exact values

User: "technical diagram: TCP flags byte (tcp[13]) bit layout with hex values and three worked sums"
→ TechnicalDiagrams intent
→ A diagram on the palette ground whose every bit label, hex value and sum is correct and legible, checked against the source before hand-over

Example 3: Comparison image

User: "make a comparison image of humans vs AI decision-making"
→ Comparisons intent
→ A hand-drawn split comparison with a clear dividing line, labeled sides and contrasting metaphors