/vision

Multi-Engine Design Direction

Shared engine selection, capability/authorization gates, dispatch, capture, attribution and degraded-mode policy: _common/MULTI_ENGINE_RECIPE.md and _common/CLI_COMPATIBILITY.md. This reference defines only the domain payload and integration rules.

Default flow for /vision multi. Run subagents in parallel — one per AVAILABLE engine — to generate divergent UX/design directions for the same brief, integrate results across two axes (concurrence + divergence), and deliver a Portfolio of complementary directions that the user selects between before downstream handoff to Muse / Palette / Flow / Forge.

Flow

SCOPE → PREFLIGHT → FAN-OUT (parallel subagents) → NORMALIZE → CLUSTER → SCORE (spectrum coverage) → GROUND (brand/persona/a11y) → SYNTHESIZE (Portfolio-only) → DELIVER

1. SCOPE

Define the design brief once. All selected subagents share the same scope:

• Operating mode (REDESIGN / NEW_PRODUCT / TREND_APPLICATION / LINEAR_RESTRAINT / SPATIAL / AI_INTERFACE)
• Product surface (marketing site / SaaS dashboard / mobile app / spatial / AI agent UI)
• Brand anchor (existing palette, typography, voice keywords, anti-keywords — if any)
• Persona target (primary persona from Cast registry if available)
• Business outcome anchor (conversion lift / task-success / retention / brand repositioning)
• Constraints (WCAG 2.2 AA mandatory, brand identity locks, technical platform limits)
• Forbidden directions (explicit anti-patterns or off-brand aesthetics)

Do NOT include the design-trend taxonomy, principle vocabulary, or aesthetic spectrum in the SCOPE — those apply during SYNTHESIZE. Let each engine reach for its own training-data priors.

2. PREFLIGHT — engine availability (Vision main context, never delegated)

Use the canonical probe from _common/MULTI_ENGINE_RECIPE.md §2. Probe codex, agy, and claude binaries with fallback paths. Apply the strict availability verdict matrix unchanged.

3. FAN-OUT — parallel subagents

Dispatch one independent task per selected, authorized engine using the shared CLI adapter for genuine parallel execution. Each subagent has independent context and generates 2–3 design directions for the brief.

Loose prompt rule (per _common/SUBAGENT.md MULTI_ENGINE): pass only Role + Target + Output format. Do NOT pass the V.A.I.R.E. rubric, 2026 trend taxonomy, design-system anti-patterns, or aesthetic vocabulary lists. Vision main context applies framework rules at SYNTHESIZE.

Required JSON output schema

{
  "engine": "codex|agy|claude",
  "directions": [
    {
      "concept_name": "Short evocative concept name (e.g., 'Quiet Industrial', 'Liquid Confidence')",
      "principles": [
        "3–5 design principles that govern this direction (each one short, declarative, testable)"
      ],
      "aesthetic_language": {
        "spectrum_position": "modernist | skeuomorphic | maximalist | minimalist | brutalist | expressive | calm | spatial | retro-futurist",
        "color_strategy": "Palette logic — saturation/temperature/contrast posture",
        "typography_strategy": "Display + body pairing, weight contrast, variable-font usage if any",
        "surface_treatment": "Flat / layered / glass / depth-token strategy",
        "imagery_strategy": "Photography vs illustration vs 3D vs generative — and tone"
      },
      "interaction_language": {
        "motion_posture": "Restrained | playful | physical | cinematic — and reason",
        "feedback_pattern": "How the system signals state changes",
        "input_modality_priority": "Pointer / touch / voice / gaze — ranked",
        "ai_disclosure_pattern": "If AI-interface: how the system shows reasoning / control / undo (required for AI_INTERFACE mode)"
      },
      "reference_style_influences": [
        "Concrete references — products, art movements, brands, OS design languages (e.g., 'Linear app restraint', 'Material 3 Expressive', 'Apple Liquid Glass', 'Swiss editorial')"
      ],
      "persona_fit": "Why this direction serves the named persona's emotional and functional needs",
      "business_outcome_link": "Which target metric this direction is hypothesized to move and how (e.g., 'Restraint reduces decision fatigue → time-on-task -15%')",
      "risk_areas": [
        "Where this direction breaks (brand drift, a11y, novelty cost, AI-trust)"
      ],
      "downstream_handoff_hint": {
        "muse": "Token themes to define (color, spacing, motion, depth)",
        "palette": "Interaction polish priorities (focus rings, hover, dead-state coverage)",
        "flow": "Motion choreography keywords",
        "forge": "Which surface to prototype first to de-risk this direction"
      }
    }
  ],
  "engine_notes": "Optional: what aesthetic bias this engine knows it brings"
}

If an engine returns free-form Markdown, ask its subagent to re-emit as JSON before integrating.

4. NORMALIZE

Parse the usable JSON outputs into a unified direction list. Tag each direction with its source engine. Preserve per-engine wording — different phrasings of the same aesthetic_language may reveal different angles on the same concept.

5. CLUSTER — same direction across engines

Group directions that share the same design concept expressed in different vocabularies. Two directions match when all of:

• Same spectrum_position (or two adjacent spectrum positions — e.g., minimalist + calm may cluster; brutalist + calm will not)
• Same motion_posture class (restrained / playful / physical / cinematic)
• Same surface_treatment class (flat / layered / glass / depth)
• Overlapping reference_style_influences (at least one shared concrete reference, OR clearly belonging to the same canonical movement)

Critical CLUSTER rule for Vision (different from Spark): Keep the same concept expressed in different idioms as separate clusters. Two engines proposing "minimalist" but one referencing Swiss editorial and the other referencing Apple HIG are different directions worth surfacing separately — the visual idiom is the deliverable. Cluster only when both the concept AND the visual idiom overlap.

Record the set of engines that produced each cluster. Preserve all per-engine wording variants of principles, aesthetic_language, and reference_style_influences for use in SYNTHESIZE.

6. SCORE — spectrum coverage + concurrence

Vision scores on two axes:

6.1 Concurrence axis (per cluster)

Vision rule (Pattern D, identical to Spark): VERIFIED-DIVERGENT directions are not automatically lower-value than UNIVERSAL. The brand-defining direction often comes from the engine with non-overlapping aesthetic priors. Surface both axes in the final Portfolio.

6.2 Spectrum coverage axis (across the full direction set)

After SCORING by concurrence, check that the surviving directions span an informative slice of the aesthetic spectrum. Reference the canonical positions:

modernist · skeuomorphic · maximalist · minimalist · brutalist · expressive · calm · spatial · retro-futurist

Coverage goals:

• Spread: at least 2 distinct spectrum_position values across the Portfolio (otherwise the user has no real choice — escalate as "low coverage, consider re-running with explicit divergence prompts")
• Contrast: at least one direction should sit clearly opposite the most-concurrent direction on the spectrum (e.g., if UNIVERSAL = minimalist, surface at least one expressive or maximalist VERIFIED-DIVERGENT if grounded)
• Mode-appropriate skew: in LINEAR_RESTRAINT mode, suppress maximalist/brutalist; in SPATIAL mode, prioritize spatial; in AI_INTERFACE mode, every direction MUST include a non-empty ai_disclosure_pattern

If spectrum coverage is poor (all surviving directions cluster within one position), flag this in the Portfolio summary and recommend a second multi run with an explicit divergence-bias prompt (e.g., "produce one direction outside the minimalist tradition").

7. GROUND — verify CANDIDATE/DIVERGENT directions (Vision main context, never delegated)

For every CANDIDATE / DIVERGENT cluster, the Vision main context must apply these grounding checks:

1. Brand-fit check — does the direction respect the stated brand anchor (palette, voice, anti-keywords)? If the direction violates a brand lock without acknowledging it as a deliberate repositioning move, mark REJECTED-BRAND-DRIFT.
2. Persona-fit check — does persona_fit reasoning hold against the named persona (ideally from Cast registry)? Generic "young professionals" framing fails. If contradicts a real Cast persona, mark REJECTED-PERSONA-MISMATCH.
3. Accessibility check — does aesthetic_language.color_strategy and surface_treatment retain WCAG 2.2 AA contrast (4.5:1 text / 3:1 UI components)? Glassmorphism, low-contrast minimalism, and brutalist high-contrast inversions are the common failure cases. If a direction implies sub-AA contrast, mark REJECTED-A11Y unless the direction itself documents the mitigation.
4. Reference-existence check — are the reference_style_influences real movements/products? Engines occasionally fabricate brand names. If hallucinated, mark REJECTED-HALLUCINATION and downgrade the cluster.
5. Outcome-link check — does business_outcome_link tie to a measurable metric, not "looks more modern"? Vague aesthetic claims fail.
6. AI-disclosure check (only when mode = AI_INTERFACE) — does interaction_language.ai_disclosure_pattern cover explainability AND user override? Silent prediction-driven UI without override is the top 2026 AI-interface failure pattern; mark REJECTED-AI-TRUST if missing.
7. Mark each as VERIFIED-DIVERGENT (keep), REJECTED-{reason} (drop), or NEEDS-INFO (ask the user).

For UNIVERSAL and LIKELY clusters, do a lightweight spot-check on a11y and AI disclosure only — three engines rarely converge on something that violates WCAG 2.2 AA simultaneously, but it does happen with low-contrast minimalism.

8. SYNTHESIZE — Portfolio-only (no Compete merge by default)

Vision differs from Spark here: design direction is a selection decision the user must make, not a synthesis the agent should perform. Vision does NOT collapse multiple directions into one "best" direction by default — that would erase the divergence that makes multi-engine valuable for design.

Portfolio Merge (Vision default)

Always emit a direction Portfolio of 3–5 complementary directions:

1. Keep all UNIVERSAL, LIKELY, and VERIFIED-DIVERGENT clusters that survived GROUND.
2. Cap at top 5 directions; if more survive, drop the lowest-business_outcome_link clarity directions first.
3. Order by: UNIVERSAL first (safe baseline), then LIKELY, then VERIFIED-DIVERGENT (breakthrough candidates). Within tier, prioritize spectrum spread.
4. For each direction, present a one-page design-direction card:
   • Concept name + engine-attribution tag + concurrence label
   • 3–5 principles (merged best wording across engines if multi-engine)
   • Aesthetic language (spectrum position, color/typography/surface/imagery)
   • Interaction language (motion, feedback, input priority, AI disclosure if applicable)
   • Reference style influences (deduped across engines)
   • Persona fit and business-outcome link
   • Risk areas
   • Downstream handoff hints (Muse / Palette / Flow / Forge — see §9)
5. Emit a single Portfolio document at docs/design/PORTFOLIO-direction-[topic]-[date].md.
6. Recommendation footer: identify Vision's lead-recommended direction with explicit concurrence/divergence reasoning (e.g., "Lead = Direction B (UNIVERSAL, lowest novelty cost); challenger = Direction D (VERIFIED-DIVERGENT, highest brand-repositioning upside)").

Compete Merge (opt-in only — multi --compete)

When the user explicitly requests a single direction (e.g., "I just want one — pick the best"), Vision falls back to a Compete merge:

1. Rank by UNIVERSAL > LIKELY > VERIFIED-DIVERGENT with spectrum-fit tiebreaker.
2. Re-mix best wording per field across engine variants of the winning cluster.
3. Emit a single direction document at docs/design/DIRECTION-[name].md with engine_concurrence front matter and an "Alternative directions considered" appendix.

Default behavior: when the user invokes plain multi without --compete, always produce a Portfolio. Vision's view is that design direction is owned by the user/team; Vision's job is to surface a curated, defensible option set.

Engine-attribution rule (both strategies)

Every shipped direction carries:

• Concurrence label: UNIVERSAL / LIKELY / VERIFIED-DIVERGENT
• Engine tag: [codex+agy+claude] (3/3) / [codex+agy] etc. (2/3) / [codex-verified] / [agy-verified] / [claude-verified] (1/3 grounded)

9. DELIVER — downstream handoff preparation

The Portfolio is the deliverable, but Vision must also prepare downstream-ready handoff packets so Muse / Palette / Flow / Forge can pick up the selected direction without re-litigating it.

For each direction in the Portfolio, include a downstream handoff stub:

The user selects one direction from the Portfolio; the selected direction's handoff stubs become the input contract for downstream agents.

Parallel Subagent Prompt Skeleton

Use the canonical spawn/capture template in _common/CLI_COMPATIBILITY.md with the JSON schema in this reference. Spawn once per selected available engine, not a fixed three. Add these domain fields; the main context owns normalization, grounding and synthesis.

Role: Generate {N=2-3} design directions for the brief below. You are one of the selected engines working independently — do not try to be exhaustive; surface what your aesthetic training-data priors suggest is most promising. Diverge from safe defaults if your training data points to a stronger angle.

Target:

• Operating mode: {REDESIGN | NEW_PRODUCT | TREND_APPLICATION | LINEAR_RESTRAINT | SPATIAL | AI_INTERFACE}
• Product surface: {marketing site | SaaS dashboard | mobile app | spatial | AI agent UI}
• Brand anchor: {existing palette, typography, voice keywords, anti-keywords — or "open"}
• Persona target: {persona name and JTBD — or "open"}
• Business outcome anchor: {conversion lift | task-success | retention | brand repositioning}
• Constraints: WCAG 2.2 AA mandatory; {brand locks}; {platform limits}
• Forbidden directions: {explicit anti-patterns or off-brand aesthetics — or "none"}

Constraints:

• Each direction names a concrete CONCEPT, not a generic adjective ("Quiet Industrial" not "Modern")
• Each direction occupies a specific spectrum_position — name it
• Each direction includes at least 3 concrete reference_style_influences (real products, movements, OS design languages)
• Each direction lists at least one risk_area
• Each direction links to a measurable business_outcome
• WCAG 2.2 AA contrast is non-negotiable; explicitly state how the color/surface strategy maintains 4.5:1 text / 3:1 UI contrast
• If mode = AI_INTERFACE: every direction MUST populate interaction_language.ai_disclosure_pattern with explainability AND user override
• Do not write implementation code — direction documents only
• Do not invent brand names or design movements; if you cite a reference, it must be a real, identifiable artifact

Degraded Modes

Use _common/MULTI_ENGINE_RECIPE.md § Engine Availability Modes and its actual-engine denominator. A healthy Claude+Codex pair is the normal dual-engine baseline, not a 2/3 degraded result.

With one usable engine, a 2–3 direction portfolio may still be offered but must be labeled single-engine; preserve user selection and brand/a11y grounding. With zero, use direction. An explicitly requested single direction routes to the ordinary recipe rather than inventing a portfolio winner.

Cross-References

• _common/MULTI_ENGINE_RECIPE.md — canonical Pattern D protocol, PREFLIGHT, FAN-OUT, scoring rubric
• _common/SUBAGENT.md §MULTI_ENGINE — base subagent dispatch and loose-prompt rules
• spark/reference/tri-engine-proposal.md — sibling Pattern D implementation (proposals); CLUSTER/SCORE pattern mirrored here
• judge/reference/tri-engine-review.md — canonical PREFLIGHT and FAN-OUT mechanics
• vision/reference/design-methodology.md — full per-mode workflow that the selected direction feeds into
• vision/reference/design-trends.md — 2026 trend catalog used at GROUND to validate spectrum coverage
• vision/reference/brand-strategy.md — brand-fit check criteria applied during GROUND
• vision/reference/agent-orchestration.md — Muse / Palette / Flow / Forge / Frame handoff contracts that downstream stubs feed
• vision/reference/output-formats.md — direction-card and Portfolio document templates

Multi-Engine Mode (SKILL.md excerpt)

Activated by the multi Recipe (or any explicit request for parallel design-direction generation / cross-engine aesthetic comparison). Mirrors Spark's Pattern D but optimizes for aesthetic-spectrum coverage and brand-defining divergence rather than feature ideation. Full algorithm, JSON schema, subagent prompt skeletons, CLUSTER/SCORE/GROUND rules, spectrum-coverage logic, downstream handoff stubs, and degraded modes → reference/tri-engine-direction.md.

• Base engine policy (2026-05): default = Claude + Codex (dual-engine, 2 spawns); agy adds a third axis (tri-engine) when AVAILABLE at PREFLIGHT — meaningful for Vision because Antigravity surfaces Material 3 Expressive / Google design-language patterns Claude and Codex under-index. See _common/MULTI_ENGINE_RECIPE.md §Base Engine Policy + §Engine Availability Modes.
• Mechanics: spawn one Agent subagent per AVAILABLE engine in one message (direction-codex + direction-claude, add direction-agy when AVAILABLE); PREFLIGHT runs in Vision main context (never delegate detection); loose prompts only (Role + Target + Output format — no rubric/trend-taxonomy/anti-pattern lists, so each engine's aesthetic priors drive divergence); subagents return JSON, main context runs NORMALIZE → CLUSTER → SCORE → GROUND → SYNTHESIZE → DELIVER.
• Concurrence scoring (Pattern D, identical to Spark/Echo[demand]): UNIVERSAL (3/3) = safe baseline (check against competitor identity to avoid same-as-everyone aesthetic); LIKELY (2/3) = strong with one dissenter; VERIFIED-DIVERGENT (1/3, grounded) = often the brand-defining direction, NOT automatically lower-value than UNIVERSAL.
• Aesthetic-spectrum coverage (Vision-specific second axis): surviving directions span ≥2 spectrum_position values (modernist · minimalist · brutalist · expressive · calm · spatial · …), with at least one opposite the most-concurrent direction. LINEAR_RESTRAINT suppresses maximalist/brutalist; SPATIAL requires spatial coverage; AI_INTERFACE requires interaction_language.ai_disclosure_pattern on every direction.
• Merge: Portfolio (default) — 3–5 complementary direction cards ordered UNIVERSAL → LIKELY → VERIFIED-DIVERGENT, each with principles / aesthetic + interaction language / references / persona fit / outcome link / risk areas / downstream handoff stubs (Muse/Palette/Flow/Forge/Frame/Prose), plus a lead-recommendation + challenger footer (docs/design/PORTFOLIO-direction-[topic]-[date].md). Compete (opt-in via multi --compete) — a single re-mixed direction (docs/design/DIRECTION-[name].md). Portfolio is default because Vision writes no code and direction is a human selection decision — collapsing to one winner erases the aesthetic breadth that makes multi-engine valuable.
• Engine-attribution tag (mandatory on every shipped direction): [codex+agy+claude] (3/3) / [codex+agy] etc. (2/3) / [codex-verified] (1/3 verified-divergent).