ArtDirection

• Nothing artistic reaches the principal or a public surface without a current passing critique from the judge panel bound to the exact bytes delivered. The panel is cross-vendor whenever two approved vendors are reachable; a one-vendor panel may still pass, and its report carries the vendor_missing flag as weaker evidence.
• Every verdict is honest: pass only when the pooled median reaches the bar and each vendor median reaches bar minus 0.5, with no UNKNOWN score or critical defect, a cap hit reported with its shortfall, and an unreachable vendor reported as unavailable rather than as a quality failure.
• Each round names ranked, located defects a producer can fix, judged against the concept-first rubric rather than adjectives.
• A figure's anatomy and mechanics are judged as a defect class: extra or missing limbs, a grip or pose that does not match the object held, or physically impossible motion is a critical defect located by image region or frame, so the critique fails it; the judges read this rule from References/Rubric.md.
• The principal's own scores are kept verbatim as labels, and judges change only through a deliberate calibration, never automatically.

Run with bun skills/ArtDirection/ prefixed to each tool path, from the config root.

Each intent file carries its tool's defaults, refusals and exit codes.

User customizations live at <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/ArtDirection/ (configRoot is $CLAUDE_CONFIG_DIR when set, else ~/.claude). When it exists, its PREFERENCES.md, configurations and resources override the defaults here; when absent, the defaults stand. First-time setup (accounts, keys, config, a proof it works) is in Setup.md.

• A brief may name "calibration": "<target>". Critique then adds that Calibrate target's Taste.json (the principal's labeled likes and dislikes) to the judge prompt and records its labels_hash in the report; a target with no profile yet is noted under outstanding. Briefs without the field produce the prompt they always did.
• Calibration labels append to ~/.claude/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/ArtDirection/labels.jsonl (fields in Intents/Calibrate.md).
• The panel is the vision vendors named by art_judges in <configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/ArtDirection/Config.md (OpenAI, Anthropic and Google have seats; a missing file or key stops the critique with an error naming both). That folder's PREFERENCES.md, when present, is appended to References/Rubric.md as install-specific judging rules, and check_set.version hashes the two together. OpenAI vision runs as codex exec -i through the codex CLI signed in to ChatGPT (Assay/Tools/OpenAIVision.ts; no API key is read); Anthropic vision runs through LIFEOS/TOOLS/Inference.ts image input (the claude CLI on the subscription), never falling back to another vendor. Panel selection (--vendors) is in Intents/Critique.md. No credentials belong in briefs or logs.
• Models come from LIFEOS/TOOLS/models.ts: CROSS_VENDOR.openaiJudge for OpenAI and the high rung of EFFORT_MODEL for Anthropic, never pinned here.
• Preferences use configRoot above; the current credential and label implementations use the home .claude location.

• Hand-over is gated by Deliver.ts because a first-draft video reached the principal with no critique on 2026-09-17, while the rule lived only in prose (the story is in the USER extension's History.md).
• The rubric leads with concept (does the artifact carry its idea as one metaphor a stranger reads at a glance), and overall follows concept most. Changed 2026-09-28 after the principal's first Calibrate scores showed the old craft-first rubric ranking the scored artifacts almost at random (the measurement is in the USER extension's History.md).

User: "iterate until it's great" on a render → Loop intent → Brief.json and Direction.md exist, critique rounds run until the artifact passes the bar, or the four-round cap is reported with its shortfall.

User: "rate the visuals" → Critique intent → Critique.json with both vendors' scores, ranked defects and a pass, fail or unavailable verdict.

User: "that one's a 6, the type is cramped" → Calibrate intent → one label appended with verdict 6 and the note in the principal's words.