stitch-ui-designer

Stitch Design Execution Orchestration

Quick Start

1. "Use Stitch to generate a store appointment tablet page; first give locally reviewable results."
2. "Only modify the top search area of the existing order page; keep the existing scope and source."
3. "Generate layout variants according to the approved quantity and verify the assets; mark missing inputs and verification status."

For designers, front-end developers, and teams maintaining this workflow. Designers provide intent and materials, developers provide engineering/tools, and the team preserves source and acceptance status during handoff.

Capability Boundaries

✅ Good at handling

• Use Stitch to generate a store appointment tablet page.
• Only modify the top search area of the existing order page.
• Generate layout variants according to the approved quantity and verify the assets.

⚠️ Requires materials

• Target project and create/edit/variant intent.
• Interface requirements, device, and existing design system.
• Available tool schema, target screens, and write scope.

❌ Inapplicable scenarios and handoff

• Only need copyable prompts → stitch-ui-prompt-architect.
• Only generate DESIGN.md → stitch-design-md.
• Implement backend APIs or publish the website → target development/deployment workflow, deliver designs and assets.

Workflow

1. Detect actual tools; when there is no MCP, output prompt-only and clearly state that nothing was generated.
2. Reuse the project, read the spec/framework contract as needed; uviewpro matching takes priority over uview.
3. Verify the applied designSystem; pass the new screen's system ID separately, do not repeat theme tokens in the prompt, keep inline tokens for legacy/prompt-only.
4. Use the architect's three-part prompt to dispatch generate_screen_from_text/edit_screens/generate_variants by intent; keep IDs as strings, and parameters according to the current schema.
5. Obtain outputComponents, project, and screen assets, verify the actual images and save the source; for interrupted non-idempotent writes, first perform three read probes, and do not resend unknown results.

Order by dependency: source verification → local artifacts → authorized external operations → verification handoff. For multiple tasks, do the current main path first; when information is missing, first output an assumption draft, then precisely list what is missing and its purpose, without using vague prompts like "please provide more context."

Security and Result Verification

Do not read unrelated account configurations, do not collect user passwords; credentials are provided only by the environment or authorized connectors. Examples use demo data only; before uploading, replace customer names, phone numbers, and order numbers with demo values, and check HTML, screenshots, and file metadata. It is forbidden to write keys, session cookies, full base64 content, or signed download URLs into reports/repositories. Unverified parameters, visual effects, and business numbers must not be fabricated; outputs must note the source, basis for decisions, what was actually executed, and what has not yet been verified.

• edit/variant selectedScreenIds come from actual queries.
• MOBILE/DESKTOP/TABLET matches the request.
• Reconcile get_project/list_screens/get_screen, and if there are no candidates, record the reason.

Device fidelity must be determined by measured values

The basis for determination is the get_screen of get_screen'stop-level deviceType、width、height；screenInstance These fields are not guaranteed to be present. After generation, read these three values first before judging; do not look only at "call succeeded."

• The provider does not guarantee that the requested device will be honored. Measured on 2026-09-14: three write paths (generate_screen_from_text、edit_screens、generate_variants) requested TABLET all returned DESKTOP 2560×2048; another MOBILE request also fell back to DESKTOP. When a fallback is hit,stopand report truthfully; do not try the other two paths instead—they will fall back in the same way and only produce more useless artifacts.
• Provider dimensions are device pixels:390×884 's viewport returns 780×1768(2×). This is scaling, not error; do not chase the 1px difference shown on the web canvas.
• Judgment rule: the device matches the request (skip this item when requesting AGNOSTIC ), and the width and height are integer multiples of the canvas, with the same multiplier horizontally and vertically. Harness's validate_screen_device based on thisfail closed; if it does not pass, it must not be marked complete, nor reported as a deliverable for that device.

Customizable: device, framework contract, generate/edit/variant mode, count, and asset directory. Value-added checks: intent routing; system token routing; interrupted write reconciliation.

FAQ

Q1: What is the main deliverable? Offline edit request: only add a "Search Orders" input before the avatar on the TABLET orders page; preserve the rest of the layout; prepare the edit_screens parameters, with 0 actual remote calls.

Q2: When should a different entry point be used instead? If you only need a copyable prompt → stitch-ui-prompt-architect; if you only generate DESIGN.md → stitch-design-md; if implementing a backend API or publishing a website → the target development/deployment process, delivering design and assets.

Q3: What happens if inputs are missing? First provide a clearly marked draft of local assumptions, and list "Needs to be supplied: target project and new/edit/variant intent; interface requirements, device, and existing design system; available tool schema, target screens, and write scope." Write operations that depend on these inputs are not executed.

Q4: How is completion determined? The selectedScreenIds for edit/variant come from an actual query; MOBILE/DESKTOP/TABLET match the request; get_project/list_screens/get_screen are reconciled, with no reason for candidate records.

Q5: How is it customized? Device, framework contract, generate/edit/variant mode, count, and asset directory; when not provided, use existing project values and state the assumptions.

Q6: Does it automatically upload, install, or go live? Only perform actions covered by the current request and existing authorization; without a remote receipt, do not claim upload success, and without runtime verification, do not claim it is live. Additional installation or scope expansion requires explaining the specific impact first.

Reference as needed

• When executing detailed mappings, APIs, or templates, read Extended Process。
• When encountering failures or abnormal input, read Anti-patterns and Gotchas。
• When edge cases, compatibility, customization, and authorization are involved, read In-depth FAQ。
• When complete input/output and validation scenarios are needed, read Local Application Example。
• The local implementation basis is the current skill's accompanying source and fixed upstream snapshot; the structure follows Agent Skills Specification. The tool's current behavior is subject to the actual schema; when not connected, do not claim that online behavior has been verified.

When to use (什么时候使用)

When the user needs to select and execute a verifiable, recoverable professional workflow for the current request , load this skill. First extract the goal, inputs, constraints, delivery format, and acceptance criteria from the request; the description summary is: orchestrate the creation, import, editing, or variant generation of Stitch screens; trigger when the user explicitly requests to actually create or modify designs with Stitch. For prompt polishing only, use stitch-ui-prompt-architect; for design documents only, use stitch-design-md..

Rules

• Read before write: first confirm the current state and actual capabilities, then perform actions that change external state.
• Least privilege: use only the files, tools, accounts, and network scope required to complete the current step.
• Evidence first: when run results, resource IDs, versions, hashes, or test outputs are missing, explicitly mark them as NOT_VERIFIED。
• Idempotency first: retain request identifiers and phase state; when results are unclear, query first rather than blindly retrying.
• Privacy and security: logs, examples, receipts, and error messages must not contain tokens, cookies, keys, or personally sensitive data.

Workflow

Step 1: Clarify intent

Confirm whether this skill matches the goal; if it is only an adjacent need, hand it to a more precise skill.

Step 2: Run prechecks

Confirm the goal, inputs, constraints, available tools, success criteria, and failure boundaries; if any key condition is unknown, stop at the read-only stage.

Step 3: Form a plan

List the tools to be called, the objects that will change, the success criteria, and the safe exit method after failure.

Step 4: Execute actions

Execute in the minimum sufficient steps, and record evidence at key state changes; retain an associable state or receipt for every external call.

Step 5: Verify delivery

Output the results, verification evidence, unfinished items, risks, and clear next steps, and state facts, inferences, and unverified items separately.

Validation checklist

• The skill trigger conditions match the user's intent, and adjacent tasks were not misrouted to this skill.
• The inputs, target objects, versions, and output locations are all clear, and no guessed values were used in place of required values.
• All write, paid, publish, or irreversible actions are within the scope of user authorization.
• The results have been verified with independent checks; "command succeeded" or "file exists" alone does not count as complete acceptance.
• The output includes actual evidence, failed/skipped items, remaining risks, and actionable next steps.

Gotchas

1. Treating the plan as the result: a document or prompt is not equal to actual execution; the actual execution level must be indicated.
2. Retrying on error: a timeout or lost response may have already produced remote state; query first before deciding whether to retry.
3. Implicit scope expansion: batch, full-scale, publish, overwrite, and paid actions are not a natural extension of ordinary reads and writes.
4. Version drift: when referencing external resources, record the version, tag, or commit; do not treat a mutable branch as release evidence.
5. Stale evidence: caches, old screenshots, and historical tests cannot prove the current environment; refresh key evidence before delivery.

Non-applicability and boundaries

Do not exceed the scope given by the user; write, paid, publish, and irreversible actions require explicit authorization. If the request requires another skill, do not copy its body text; hand off by skill name and preserve the current task context.

Progressive disclosure

• When you need to determine inputs/outputs, states, and authorization points, read references/workflow-contract.md。
• When you need to perform a pre-delivery self-check, read references/validation-checklist.md。
• When encountering timeouts, partial success, or recovery scenarios, read references/error-recovery.md。
• For first run, denial of privilege escalation, and failure recovery, refer respectively to examples/happy-path.md、examples/boundary-refusal.md、examples/failure-recovery.md。