Design loop

Design Loop

Turns a project directory into a customized, multi-loop design exploration prompt. Does not execute the loop — produces the prompt that a separate loop runner consumes.

When to use

• User points at an existing codebase and asks to improve, refine, or polish the UI
• User wants a structured design exploration, not a single mockup
• User mentions "design loop", "explore directions", "modernize", "visual refresh", "UI audit"

When NOT to use

• Greenfield design with no existing code — use the brainstorming skill. This skill exits early if discover.py finds no UI surface.
• Single-component tweaks ("change this button color") — just make the change.
• CSS debugging or pixel-perfect bug fixes — those are surgical, not exploratory.
• Pure backend projects with no UI — exit early and say so.

Workflow

1. Confirm the target path

If the user didn't supply one, ask. If the path doesn't exist or isn't a directory, stop and surface that before running anything else.

2. Run discovery

mkdir -p <target_path>/design-exploration
python "<SKILL_DIR>/scripts/discover.py" <target_path> \
  > <target_path>/design-exploration/context.json

Handle two early-exit cases based on what discover.py emits:

• No UI surface detected (no component dirs, no style files, no UI framework, project_type == "unknown" with zero domain hits): stop. Tell the user this looks greenfield or backend-only and suggest the brainstorming skill. Do not generate a prompt.
• Multiple UI surfaces detected (monorepo with e.g. apps/admin + apps/marketing): discover.py emits a ui_surfaces array. Ask the user which surface to target, then re-run with --surface apps/admin passed to generate_prompt.py.

3. Surface discovery decisions

Before generating the prompt, print these fields from context.json so the user can override:

• project_type
• ui_framework
• style_system
• First 5 references that would be auto-picked for this project_type
• component_dirs and any existing screenshots/mockups
• data_version of the references/dimensions libraries

Ask: "Override any of these with --type, --refs, --surface, --hero, or --runner?"

4. Generate the prompt

python "<SKILL_DIR>/scripts/generate_prompt.py" \
  <target_path>/design-exploration/context.json \
  > <target_path>/design-exploration/LOOP_PROMPT.md

5. Validate the output contract

python "<SKILL_DIR>/scripts/validate_prompt.py" \
  <target_path>/design-exploration/LOOP_PROMPT.md

If validation fails, the template or generator regressed — fix before handing off. Do not ship a prompt that doesn't pass validation.

6. Hand off

Print the full path to LOOP_PROMPT.md and the runner invocation. The default is /loop, but any exploration runner works — the prompt itself is runner-agnostic.

Prompt ready: <target_path>/design-exploration/LOOP_PROMPT.md
Suggested invocation: /loop 10m  (6 loops)

If the user passed --runner to generate_prompt.py, echo that invocation instead.

Overriding discovery

Pass hints through generate_prompt.py to skip or correct discovery:

Flag	Purpose	Default
--type	saas-dashboard | marketing | dev-tool | internal-ops | ecommerce | creative | game | docs	auto-detected
--refs	Comma-separated reference override, e.g. linear,vercel,stripe	auto-picked from library
--surface	UI surface path in a monorepo, e.g. apps/admin	first detected
--loops	Loop count	6
--duration	Per-loop duration	10m
--hero	Final hero mockup target	type-appropriate default
--runner	Runner invocation embedded in handoff	/loop
--min-viewport	Minimum viewport spec	1280x800

Example:

python generate_prompt.py context.json \
  --type creative --refs figma,ableton,arc --loops 8 --runner /loop

Output contract

Every generated prompt MUST include the items below, and validate_prompt.py asserts each one. If you add a required section, update the validator in the same commit.

1. Mission statement naming the project and its detected type
2. Reference set (5 products, studied not copied)
3. Per-loop deliverables requirement (minimum one of five artifact types)
4. Exploration dimensions scoped to the project type
5. Detected stack constraints (framework, style system, component dirs)
6. Loop-N convergence clause producing manifesto.md + hero mockup
7. Data library version tag (so stale prompts are spottable later)

Files

design-loop/
├── SKILL.md
├── scripts/
│   ├── discover.py           # scans target, emits context.json
│   ├── generate_prompt.py    # fills template from context
│   └── validate_prompt.py    # asserts output contract
├── templates/
│   └── loop-prompt.md.tmpl   # parameterized exploration prompt
├── data/
│   ├── references.json       # project-type → reference products (versioned)
│   ├── dimensions.json       # project-type → exploration dimensions (versioned)
│   └── README.md             # schema + allowed project types
├── fixtures/
│   ├── nextjs-dashboard/     # sample + expected outputs
│   ├── streamlit-app/
│   └── static-site/
└── tests/
    └── test_pipeline.py      # end-to-end fixture diff test

Tuning

The reference and dimensions libraries are the highest-leverage places to improve output quality. When a generated prompt produces weak design directions, first check whether the references and dimensions matched the project. Edit data/references.json and data/dimensions.json rather than rewriting the template.

Both data files carry a "version" field (e.g. "2026.04"). Bump it on non-trivial edits. generate_prompt.py writes the version into the prompt header and warns when a context.json was produced against an older library version than the current one.

See data/README.md for the full schema, allowed project_type values, and field definitions — edit that doc if you extend the libraries.

Testing

Run pytest tests/ after any change to discover.py, generate_prompt.py, the template, or the data files. Fixtures under fixtures/ are synthetic project skeletons with pinned expected outputs; the test diffs the generated context.json and LOOP_PROMPT.md against the expected files and fails loud on drift.

When a fixture's expected output changes legitimately (e.g. you bumped the references library), regenerate and commit the updated expected files in the same PR.