Sumi Orchestrator

Sumi Orchestrator

Sumi holds 43 skills and 37 commands. Most real requests need three to six skills in a particular order, and running them in the wrong order wastes work — you cannot critique a flow before you know who it is for, and you cannot set metrics before you know what good looks like.

This skill is the router. Identify the pipeline, run its stages in order, honour the gate, hand off the named artifact, and name the command that produces it.

Two bands: stages and depth

Not every skill is a stage. A stage is a decision that gates the next decision. A depth skill is a library you open once you already know what you are building — 500+ layout blocks, 1000+ microcopy templates, 200+ animation recipes, 500+ palettes. Loading a library before you have a decision to make just burns context.

So each pipeline below lists its core stages in order, then the depth skills to pull in inside a stage when the work gets that specific. Keep core stages at six or fewer. If a pipeline seems to need more, one of them is depth.

Routing

Match the user's request to a pipeline. If the request is vague ("help me with my app"), ask the qualifying questions below before routing.

| The user wants to… | Pipeline | Command | |---|---|---| | Find what's wrong with something that exists | Evaluate | /audit, /roast | | Fix AI-generated UI that works but looks wrong | Fix | /fix | | Design something that doesn't exist yet | Create | /style → /screen | | Build or ship an actual component | Implement | /component | | Compose a whole page | Compose | /page, /layout | | Generate design assets with AI | Generate | /generate | | Start or scale a design system | Systematize | /tokens | | Move a design into code | Handoff | /figma | | Understand why people aren't finishing | Convert | — | | Ship to new countries or languages | Localize | — | | Design an AI or agent feature | AI Surface | /ai-audit | | Know whether any of it worked | Measure | /measure |

Qualifying questions

Ask at most three, and only what the request hasn't already answered:

1. Does it exist yet? Nothing built → Create. Something built → Evaluate.

Built by an AI and looks off → Fix.

1. What platform? Web/desktop, mobile, native, or ambient/voice/spatial. This

selects the platform skill layered into every pipeline.

1. What decision hangs on this? Ship/no-ship, roadmap priority, or a specific

fix. This sets the depth — a ship gate needs the full pipeline; a specific fix usually needs one skill.

Pipelines

Each stage names the skill to invoke. Gate is the condition to satisfy before moving on — if it fails, stay in the stage or step back, don't push forward.

Evaluate — something exists and may be broken

1. nng-ux-heuristics — heuristic pass, severity 0–4 on every finding
2. cognitive-psychology-ux — cognitive load, decision architecture, attention
3. accessibility-inclusive-design — WCAG 2.2, ARIA, keyboard, contrast
4. ui-pattern-intelligence — anti-patterns, pattern matching, AI-slop detection
5. ux-ethics-content-strategy — dark patterns, manipulation, interface copy
6. design-critique-case-studies — structure the findings into a critique

Depth: visual-design-mastery when the complaint is "it looks cheap"; micro-copy-intelligence when the findings are about strings; conversion-optimization-patterns for a funnel; data-visualization-mastery for charts; platform-visual-standards for native. Gate: every finding carries a severity rating and a location. Unrated findings are opinions, and opinions don't survive a prioritization meeting. Artifact: severity-ranked defect list.

Fix — AI-generated UI that runs but looks wrong

1. ui-pattern-intelligence — detect the slop patterns and name them
2. visual-design-mastery — score it and cite the canonical rules it breaks
3. ui-visual-design-system — replace arbitrary values with a system
4. component-patterns-code — rewrite it properly
5. accessibility-inclusive-design — the states and labels the generator skipped
6. micro-copy-intelligence — replace placeholder and robotic strings

Depth: color-palette-library, typography-pairing-recipes, shadow-elevation-density, animation-recipe-library. Gate: no hardcoded value that a token could carry, and every state the generator skipped now exists. Report the before/after score delta. Artifact: corrected code plus the score delta that justifies it.

Create — designing something new

1. ux-research-methods — what do we actually know about these users
2. cognitive-psychology-ux — the constraints the design must respect
3. sector-style-intelligence — what credible looks like in this category
4. screen-flow-patterns — which screens exist and what connects them
5. ui-visual-design-system — type, color, hierarchy, spacing
6. interaction-motion-design — how it moves and responds
7. performance-states-patterns — empty, loading, error, offline, success
8. accessibility-inclusive-design — before it's built, not after

Depth: navigation-pattern-encyclopedia, form-design-encyclopedia, micro-copy-intelligence, ui-pattern-intelligence. Short form: stages 3, 4, 5, 8 when the decision doesn't need the full run. This is the one pipeline that exceeds six stages — say so before starting it, and offer the short form. Gate: every screen has all its states specified. A design that only shows the happy path is a mockup, not a design. Artifact: design specification with a state matrix.

Implement — building the real thing

1. design-systems-architecture — which tokens and components already exist
2. ui-visual-design-system — visual spec for anything new
3. component-patterns-code — production code, full state matrix, ARIA
4. micro-copy-intelligence — every component ships strings; robotic labels are

the loudest slop tell and the cheapest thing to fix

1. accessibility-inclusive-design — verify against the built artifact
2. performance-states-patterns — perceived performance and state handling

Depth, routed by component type: form → form-design-encyclopedia; nav → navigation-pattern-encyclopedia; chart or dense table → data-visualization-mastery; media → image-media-patterns; icon → icon-illustration-systems; motion → animation-recipe-library; breakpoints → responsive-block-patterns; native → platform-visual-standards. Gate: keyboard-navigable, screen-reader-labelled, every state rendered. Artifact: component code plus its state and a11y checklist.

Compose — building a whole page

1. sector-style-intelligence — the register the page has to hit
2. page-composition-engine — block order, spacing rhythm, visual pacing
3. layout-block-intelligence — the individual sections
4. micro-copy-intelligence — headlines and body that carry the argument
5. responsive-block-patterns — how each block transforms on small screens
6. performance-states-patterns — loading behaviour and perceived speed

Depth: image-media-patterns, animation-recipe-library, conversion-optimization-patterns, typography-pairing-recipes. Gate: every block has a job in the page's narrative and a mobile form. No block is decoration. Artifact: page composition with block order and responsive behaviour.

Generate — AI-generated design assets

1. sector-style-intelligence — the direction to generate toward
2. ai-design-generation — the MCP models, prompts, and pipelines
3. visual-design-mastery — score what came back
4. ui-pattern-intelligence — check it against real patterns, not plausible ones
5. component-patterns-code — turn the accepted output into real code

Depth: icon-illustration-systems, image-media-patterns, color-palette-library. Gate: nothing leaves this pipeline without a visual-quality score and a named verdict — accept, regenerate, or hand-build. Generation's failure mode is exactly the one Sumi exists to fix, so the gate is the point of the pipeline. Artifact: generated assets plus the score and verdict for each.

Systematize — starting or scaling a design system

1. design-systems-architecture — token tiers, governance, versioning
2. design-token-presets — start from a vetted vertical preset, not a blank file
3. color-palette-library — the scale the semantic tokens alias
4. typography-pairing-recipes — the type scale
5. component-patterns-code — reference implementations
6. figma-design-tool-workflows — keep design and code in sync

Depth: shadow-elevation-density, ui-visual-design-system, visual-design-mastery, platform-visual-standards. Gate: tokens are a single source of truth in W3C DTCG format, one change propagates everywhere without hand-editing, and every component consumes a semantic alias rather than a primitive directly. Artifact: token file, component inventory, governance model.

Handoff — design to code

1. figma-design-tool-workflows — Auto Layout, variables, Dev Mode, MCP
2. design-systems-architecture — map Figma variables onto real tokens
3. component-patterns-code — generate the platform code
4. responsive-block-patterns — a frame set is fixed-width; responsive rules are

the part that never survives handoff unless someone writes them down

Gate: generated code consumes tokens, not hardcoded values, and every breakpoint transformation is specified rather than inferred. Artifact: platform code traceable back to the source design.

Convert — people arrive and don't finish

1. conversion-optimization-patterns — funnel friction, CTA, pricing, checkout
2. cognitive-psychology-ux — the decision architecture underneath it
3. form-design-encyclopedia — where forms are the drop-off
4. micro-copy-intelligence — the strings that carry or kill the decision
5. ux-ethics-content-strategy — the ethics check, as a gate
6. ux-metrics-measurement — how you would know it worked

Gate: every recommendation names the funnel step it affects and the metric that would move — and clears stage 5. This is the pipeline where Sumi could most easily be turned into a manipulation engine, so the dark-pattern check is a gate, not a courtesy. A lift that depends on deceiving someone is not a finding, it is a liability. Artifact: ranked funnel-friction list with a test plan.

Localize — new markets

1. cross-cultural-i18n-ux — RTL, CJK, expansion, locale formats, color meaning
2. typography-pairing-recipes — CJK, Arabic, and Cyrillic coverage is a font

problem before it is a layout problem

1. ui-visual-design-system — logical properties, flexible containers
2. accessibility-inclusive-design — language attributes, reading order

Depth: responsive-block-patterns, micro-copy-intelligence. Gate: layout survives 40% text expansion and RTL mirroring. Artifact: localization readiness report.

AI Surface — designing an AI or agent feature

1. agentic-ai-generative-ux — agent patterns, confidence, guardrails
2. ai-spatial-voice-ux — conversational, voice, multimodal patterns
3. micro-copy-intelligence — an agent's trustworthiness is almost entirely its

strings

1. ux-ethics-content-strategy — consent, control, accountability
2. cognitive-psychology-ux — trust calibration and automation bias

Depth: performance-states-patterns for streaming and latency, ambient-calm-zero-ui for screenless surfaces. Gate: the user can tell what the AI did, why, and how to undo it. Artifact: AI interaction spec with trust and recovery affordances.

Measure — did it work

1. ux-metrics-measurement — HEART, SUS, task success, statistical validity
2. ux-research-methods — the study design that produces those numbers
3. data-visualization-mastery — every one of these plans ends in a dashboard,

and an unreadable dashboard is an unused one

Depth: conversion-optimization-patterns. Gate: every metric has a baseline. A metric without a baseline cannot show improvement. Artifact: measurement plan with baselines and instrumentation.

Layers

Platform

Layer the matching platform skill into whichever pipeline is running:

• Mobile or responsive → mobile-ux-design + platform-visual-standards
• Desktop, enterprise, dashboards, data-dense → desktop-app-design +

data-visualization-mastery

• Ambient, wearable, automotive, screenless → ambient-calm-zero-ui
• Native iOS or Android specifically → platform-visual-standards

Process

If the user asks what the process is called, wants a design sprint, or needs a Double Diamond framing for stakeholders → design-process-methods.

Packaging

If the artifact is leaving the team — client, executive, portfolio — package it with business-design-templates.

Memory

Any pipeline spanning more than one command relies on .sumi/ to carry decisions between them. design-memory holds that contract — the canonical schema, which command owns which subtree, and the merge rules. Load it whenever a stage reads or writes design memory, which in practice is every Create, Compose, Systematize and Fix run.

How this relates to the commands

• Commands do not invoke pipelines. A command has a fixed output contract.

Silently running a six-stage pipeline behind it would break that contract and spend context the user didn't ask for. Each command's own "Next Step" footer surfaces the pipeline one step at a time, which is the right pace.

• This skill does not invoke commands. It names skills, because skills carry

the knowledge. But every pipeline above names the command that produces its artifact, and you should hand off to it by name.

• The chain is /start → orchestrator → pipeline → command. This is why

/start never has to list 37 commands.

• Convert and Localize have no command yet. Run them as pipelines and say so.

Operating rules

• Announce the pipeline before running it. The user should know which stages

are coming and roughly what each costs them in time.

• One stage at a time. Each skill's content stays in context once loaded, so

running six at once is both slower and worse than running them in sequence. This is also why depth skills are opened inside a stage, not up front.

• Report gate failures. "Stage 3 failed the gate — 12 findings have no

severity rating" is useful. Silently continuing is not.

• Short-circuit narrow requests. "What's a good tap target size?" is one

skill, not a pipeline. Don't run process for its own sake.

• Stop when the decision is answerable. The pipeline serves the decision, not

the other way around.

Cross-references

Each pipeline's stages have their own reference files, listed inside each skill. Load references only when the stage's SKILL.md points to them — they are large and they persist in context once read.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: phazurlabs
• Source: phazurlabs/sumi
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.