Art Direction

Art-Direction — one committed direction, written as law, for any medium

art-direction fills the seam between ideate and build-loop: ideate decides whether feel is load-bearing, build-loop iterates against a design bar — this skill produces the bar. It takes anything visual (a product, a site, a game, a poster, a chart system, a brand, a film) through a disciplined funnel ending in one durable artifact: ART_DIRECTION.md — a short constitutional brief that pre-answers every later visual decision, and the only file handed downstream. A working log, ART_DIRECTION_NOTES.md (diagnosis → panel verdict → critic trail → honest residual), travels beside it but never enters a critic's inputs — critics review blind.

ideate → CONCEPT_BRIEF.md (feel load-bearing?) → art-direction → ART_DIRECTION.md → prompt-pack / build-loop (the design bar + fidelity critic) (autopilot's fixed pipeline doesn't route here on its own — when its run marks feel load-bearing, this skill's proactive trigger is the entry point, between ideate and prompt-pack.)

The method is reverse-engineered from a real rescue: a fully built product whose visual layer scored 79/100 on effect-craft and still looked bad — every individual effect competent, no direction governing any of it — then was rebuilt the same day from a committed one-page brief into work an independent different-model critic called "genuinely calm, premium, cohesive." The delta was not more effects or more critique. It was a constitution.

Why this exists (the problem it solves)

Left ungoverned, visual work with an AI (or a team) fails in one canonical way — the pile of effects — assembled from five predictable parts:

• Critic-accretion. Each review round adds: "a critic said add bloom, so bloom was added; add an aurora, so an aurora was added." Round after helpful round: six contradictory lighting models. The critic is the failure engine when the critic is allowed to prescribe additions.
• Scores reward the wrong axis. The proven failure's 79/100 came from a blind, different-model critic scoring holistic effect-craft — it called the pile "premium." Critic independence didn't save it; what was scored did. Hence the fixed cohesion axes.
• No quiet baseline. Everything sits at 7/10 loudness, so the one moment that matters has no contrast headroom — and gets "fixed" with more glow, which raises the floor further.
• Decoration out-competes structure. The theme wins the contrast war; users can't find the button but love the flourish.
• Skin, not test. A theme that decorates but never decides — a paper texture pasted behind the same glassy buttons. A direction with veto power over nothing is a skin by definition.

Each of these is prevented structurally (see Forcing functions and the critic charter), not by exhortation.

The prime directive

Your first deliverable is not code, not artwork, not a moodboard. It is a decision. Commit to ONE direction, write it down as law, and build only what that document authorizes. If you skip the page and start executing, you will reproduce the exact failure this skill exists to prevent.

When to use this

Strong triggers — invoke without asking:

• "Art direction / design direction / visual identity / design language for X"
• "Make it cohesive" / "commit to a style" / "it looks tacky / generic / AI-generated"
• "Redesign the look" / "style guide" / "art bible" / "this design got worse the more we iterated"

Softer triggers — invoke proactively:

• ideate marked feel load-bearing and the brief's design direction is still a vibe-phrase, not a law — author the real direction before prompt-pack/build-loop run.
• A design review loop is visibly accreting effects with no governing document to check fidelity against.

Do NOT use this for:

• Execution craft under a direction that already exists → compose frontend-design (web), dataviz (charts) — under the committed brief, which supersedes their own pick-a-style steps.
• Deciding WHAT to build or whether it's worth building → ideate.
• Driving a build to done → build-loop (this skill hands it the bar; it doesn't run the build).
• A plain utility where design isn't load-bearing → plain-but-clear is correct; don't gold-plate.

Routing tie-breaker: art-direction answers "what single direction governs how this looks and feels — and what does it forbid?" — not "what should we build?" (ideate), "does it work?" (build-loop), or "how do I execute this well?" (frontend-design/dataviz).

What it produces — and its honest bounds (read this before you trust it)

Produces: one committed ART_DIRECTION.md (a short page of quantified law); a saved diagnosis + panel verdict; a north-star proof (representative work executed to final quality under the law); the critique protocol downstream loops enforce; and an honest residual.

Bounds, stated up front:

• Cohesion is checkable; beauty is not. The skill guarantees a governed, coherent, defensible direction — the property whose absence makes work read as tacky. Whether the committed direction is great is taste, and taste stays human-gated: interactively, the commit is offered for sign-off; autonomously, the direction choice is flagged as the run's single biggest taste call in the residual. Never claim a model-validated direction is designer-approved.
• The direction is chosen by structural fit (which candidate cures the diagnosed failure by construction), not by beauty contest. That's the honest, defensible criterion — it is also why the result can be good without anyone's taste being proven.
• Market appeal is out of scope. A cohesive look ≠ a look people want. That's the human/market handoff, as everywhere in this family.

The method (run in order)

Detailed moves per phase live in references/process-guide.md. Provenance, honestly: Phases 0 (autopsy mode), 1–4, 6, and 7 mirror the proven run; the pre-mortem mode, the detailed panel protocol, Phase 5's north-star proof, and §11's amendment log are hardening added in generalization — adopted by design, not by precedent. In brief:

Phase 0 — Diagnose before any aesthetics. Two modes: Autopsy (something exists and fails — render or gather it, name the root cause forensically) or Pre-mortem (greenfield — name the medium's ambient defaults you'd sleepwalk into, the hero moment the design must make unmissable, and the hardest legibility constraint). Also inventory the givens — locked assets, parent brand, format specs the direction must live inside. Pressure-test the diagnosis with an independent read, then write it into ART_DIRECTION_NOTES.md. The output is a named enemy the direction must cure. No diagnosis → no direction.

Phase 1 — Reference foundation. Gather 5–8 references that are cohesive, and compress each to its one governing idea ("every frame should be printable"; "weather is the light engine") — internalize the why, never copy the what.

Phase 2 — Author 3 expensive candidates. Each fully committed on paper: a named world, material, palette values, depth model, type plan, emphasis-budget sketch, its own membership test, compliance with the givens — and an explicit claim of which diagnosed failure it cures by construction. Cheap one-line candidates make the panel vote meaningless and the commit weak.

Phase 3 — Independent panel. 3 independent judges (different model or genuinely fresh contexts), blind to each other, given only the diagnosis + givens + candidates; a candidate that violates a given is disqualified. Selection criterion: structural fit to the diagnosis, not prettiest. Full protocol (including the split-vote path and the no-isolation degradation rule) in references/critique-protocol.md; verdict saved to the notes file.

Phase 4 — COMMIT. Write ART_DIRECTION.md before any execution. This is the gate. A short page of law (skeleton below; template in references/brief-template.md). Offer it for sign-off when the human can respond; otherwise proceed as "Committed (pending sign-off)" and flag it in the residual. Update the CONCEPT_BRIEF's design field to point at the committed law so downstream skills discover it.

Phase 5 — North-star proof. Execute representative work to final quality under the law — ONE piece for a single-surface subject; for a series, template, or multi-surface subject, a mini-sheet of 2–3 instances including the harshest-constraint surface. If the law fails its first contact, amend it deliberately (recorded), don't silently drift. The proof becomes the grading reference for everything after.

Phase 6 — Governed execution. Hand off to build-loop (the brief is its design bar — and the handoff must embed the critic charter, the ten axes, and §11's removed list into the design acceptance criteria, superseding its critic's own defaults) or execute inline under the same loop. Verify against the rendered artifact, state by state, climax in motion — per-medium substrates in references/media-notes.md.

Phase 7 — Honest residual + the enforcement stack. Close the notes file with the three-part residual (what this environment can't verify · out-of-scope flags · taste calls a human should revisit), then replicate the lock wherever future work enters: project instructions, design tokens or locked master assets as the single source of truth, the updated CONCEPT_BRIEF, stale-doc quarantine, a memory note, and an embedded copy of the critic charter.

The brief is a constitution, not a moodboard

Default-deny: "This document governs every [pixel/frame/spread/second]. If an element is not authorized below, it does not exist." A status header (§0) plus eleven numbered sections, every law carrying a number or an enumerable — full template with law-phrasing patterns in references/brief-template.md:

§0 status, authority & givens (+ the named enemy) · §1 the single unifying idea + the membership test · §2 the named material (behavior, seams, forbidden materials) · §3 palette as law (roles, proportions, the scarcity economy) · §4 one depth model (counted cues) · §5 type as citizens (exclusive roles, mechanical justification) · §6 the emphasis budget (inert baseline, closed ledger, everything settles back) · §7 restraint (the do-NOT list, counted budgets, the designed zero state) · §8 identity & legibility defense (multi-channel identity; structure out-contrasts decoration; the pre-refuted "volume fix") · §9 the climactic moment (numerically choreographed, native to the world, then spent — or the signature gesture, for subjects with no single climax) · §10 medium mechanics (references/media-notes.md) · §11 provenance (traceability, removed/rejected list, kept list, amendment log).

The length law: the brief stays re-readable in one sitting, before any work session — ≤ ~160 lines (the proven instance is 161). A bible nobody re-reads governs nothing — depth goes in the references and the notes file, never the law.

The critic charter (fidelity, never addition)

Every critique pass — panel, build-loop's different-model critic, your own — runs under this charter. It is the single line that kills critic-accretion:

• Allowed: fidelity violations citing the authorizing section ("§4 permits two depth cues; this card invents a third"); regressions against the removed list; subtraction proposals; diagnoses of what isn't working, stated as problems ("the CTA reads as a foreign material"). Every finding carries the family severity tier (Blocker/High/Medium/Low) so it merges straight into build-loop's defect ledger.
• Forbidden: prescribing any new effect, element, or mechanism. A genuinely new idea enters only as a proposed amendment to the brief — a deliberate, recorded act by the direction's owner, never a critique note.
• Critics score the fixed cohesion axes (references/critique-protocol.md) against the rendered artifact, every pass, and assert "no removed-list item resurfaced." Axis scores are the trajectory signal, never the gate: act on agreed structural defects; never chase the number.

Forcing functions (the skill must make these happen)

• The test must have vetoed something. §1's membership question is real only if the do-NOT list and the rejected list are its receipts. An empty do-NOT list means you wrote a skin.
• Every law is quantified. "Mostly neutral" is unenforceable; "~80% of every frame is neutral ground" is. A number or an enumerable in every clause.
• Scarcity: full intensity is earned and spent. Peak chroma/loudness exists only at the climactic moment, as a lift of the family's own values — never new colors, never white-hot — then returns to baseline.
• Hero-nativity. Choose the world whose physics produce the product's hero moment by construction — the climax should be what the world's physics naturally do, not a thing added to the theme.
• The seam law. One material spans the medium's real subsystem boundaries (chrome/content, page/app, breakpoints, panels, scenes — and, for multi-surface subjects, the surfaces themselves). Seams are where directions die.
• Identity is never single-channel. Hue + shape/pattern/position, so the quiet survives grayscale, CVD, thumbnails, one-color print — and legibility complaints get a redundant channel, never more volume.

Pitfalls to avoid (each maps to a real failure)

• Critic-accretion → the charter: critics diagnose and subtract; additions only by recorded amendment.
• Chasing the score → the 79/100 lesson: an independent critic validated the pile because it scored the wrong axes. Fixed cohesion axes; act on agreed structural defects only.
• Skin, not test → if the direction has never forbidden anything, it decides nothing. Lint for the empty do-NOT list.
• The volume fix → "too muted → brighten it" destroys the load-bearing quiet; §8 pre-refutes it. Add a channel instead.
• Ignoring the givens → a gorgeous direction the parent brand (or the footage, or the format spec) forbids is a dead direction. Inventory givens in Phase 0; disqualify violators at the panel.
• Drift by twenty reasonable steps → tokens as single source, provenance per effect, the same axes scored every pass, amendments recorded.
• The bloated bible → a brief too long to re-read governs nothing. ≤ ~160 lines of law; depth in references.
• Verification theater → cohesion asserted from intentions or code, or a "panel" of simulated judges in one context. Score only the rendered artifact; label a self-judged verdict SELF-JUDGED; a check you couldn't run is reported NOT RUN, never passed.
• Gold-plating a utility → if feel isn't load-bearing, plain-but-clear is correct; say so and stop.

Scale heuristics

The skill scales down gracefully — but the discipline is constant: a named enemy, one committed law, critics who may never add, and an honest residual of what only a human can judge.