Agent Skills: Authoring Stitch Prompts

Authoring Stitch Prompts

Quick Start

1. Collect context – accept natural language, specs, or referenced files describing the screen/app. 1.5. Discover design context (optional) – check for design-intent/:
   • If exists: Extract Project Type, Design System from design-intent/memory/constitution.md
   • If not found: Scan codebase for framework hints (package.json)
   • Falls back gracefully to standalone mode
   • See WORKFLOW.md for details.
2. Parse essentials – identify app type, screen focus, layout elements, and visual cues.
3. Detect split points – analyze if input contains multiple screens or distinct intents (>2). Apply smart defaults: split if >2 screens/intents, else combine. Users can request regeneration with different approach.
4. Filter aggressively – strip ALL non-UI concerns (backend, auth, APIs, caching, error handling, performance metrics, code-level specs). Focus EXCLUSIVELY on visual layout, components, colors, typography, spacing, and interaction patterns.
5. Condense – rewrite into one atomic Stitch directive using "Design/Create/Add…" phrasing.
6. Structure output – follow the Stitch prompt template (directive sentence → bullet list → 3–6 style cues → constraints). If design context was discovered, inject project-appropriate style cues. Do NOT use multi-section headings.
7. Validate – ensure UI nouns are present, word count <250, NO technical implementation terms, and format matches EXAMPLES.md structure before returning the prompt.

Use this Skill whenever users need Stitch-ready wording, prompt refinements, or style-consistent rewrites.

File Output

Prompts are saved to feature-based directories under design-intent/google-stitch/{feature}/:

• Feature name derived from screen/page purpose (kebab-case, semantic, concise)
• Prompt files: prompt-v{version}.md (auto-incremented from existing versions)
• Pre-created subdirectories: exports/ (Stitch outputs) and wireframes/ (reference mockups)
• File composition: <!-- Layout: {Name} --> header, then --- separators between component sections (<!-- Component: {Name} -->)
• 6-prompt Stitch limit: If >6 prompts, split into prompt-v{version}-part{N}.md files (max 6 per part)
• Copy-paste ready: Entire file works directly in Stitch interface

See WORKFLOW.md for detailed file composition rules, REFERENCE.md for Stitch best practices, and EXAMPLES.md for worked examples (Examples 14–16 cover multi-component and split scenarios).

Input Types

Accepted

• Natural-language descriptions (single screen or short flows)
• Markdown/YAML/JSON specs (/specs/dashboard.md)
• Revision directives ("move KPI cards above chart", "convert to French", "change button to green")
• References to uploaded wireframes or images
• Language conversion requests ("switch to Spanish", "German version")
• Structured input from /prompt command (see below)

Structured Input (from /prompt command)

When invoked via /prompt, the skill may receive pre-parsed preferences (Brief, Components, Style, Structure). Parse structured fields and apply style mappings before proceeding. See WORKFLOW.md for field handling and style mapping tables.

All input detail levels are valid — Stitch infers patterns from minimal descriptions. Use adjectives to convey vibe when details are sparse ("vibrant fitness app", "minimal meditation app").

Workflow Overview

High-level loop: parse → condense → format → validate. Detailed branching logic, including cue extraction and revision handling, lives in WORKFLOW.md.

Output Structure

Prompts must follow the Stitch-friendly template:

• One-sentence description of the app/screen + primary intent.
• Bullet list (3–6 items) covering layout, components, or flows.
• Visual style cues (palette, typography, density, tone).
• Optional behavior/constraint reminders (responsiveness, export format).

Reference templates/authoring-stitch-prompts-template.md for wording patterns and templates/layout-prompt-template.md for layout/foundation prompts.

Examples

Representative before/after samples (SaaS dashboard, banking app, iterative edits, spec conversions) are in EXAMPLES.md. Use them to mirror tone and formatting; keep this file lean by not re-embedding the full transcripts here.

Design Context Integration

When design-intent/ exists in the project, the skill enhances style cues with project context:

• Project Type influences tone (e.g., "enterprise-grade" for Enterprise, "friendly, approachable" for Consumer)
• Design System names appear in style cues (e.g., "Fluent UI styling", "Material Design patterns")

The skill does NOT inject specific tokens (hex colors, spacing values)—only high-level descriptors that help Stitch generate contextually appropriate designs.

Fallback behavior: If design-intent/ is not found, the skill works standalone using default style cues.

See WORKFLOW.md for discovery logic and WORKFLOW.md for injection rules.

MCP Integration (Optional)

When @_davideast/stitch-mcp is configured, prompts can be sent directly to Stitch for generation after authoring.

With MCP: Author prompt -> Generate screens -> Fetch images/code Without MCP: Author prompt -> Copy to Stitch manually

After authoring, offer: "Stitch MCP is available. Generate screens now? [Yes / No / Just save prompts]"

If accepted, invoke the generating-stitch-screens skill with the authored prompt file.

Common Issues

• Prompts too verbose – Re-run formatting with the template and trim narration. See TROUBLESHOOTING.md.
• Missing style cues – Derive palette/typography keywords from user input or prior session context before finalizing. See TROUBLESHOOTING.md.
• Multi-goal briefs – Split into multiple prompts; re-emphasize Stitch's atomic focus. See TROUBLESHOOTING.md.

Reference Files

For advanced usage:

• REFERENCE.md — Overview of Stitch best practices
• EXAMPLES.md — Sample transformations
• WORKFLOW.md — Detailed processing loop
• TROUBLESHOOTING.md — Error-handling guidance
• templates/authoring-stitch-prompts-template.md — Output format template