Daisy AI Agents Skill

Drive Daisy directly from your AI coding agent. Ask your agent "design me a settings screen" or "mock up onboarding for my fitness app" and watch screens materialize in your daisy project — the agent reads back each screen's rendered HTML and can render them to PNG/JPEG images to drop into a chat or PR.

Installation

Browse and install interactively:

npx skills add daisy-now/agent-skills

What this skill does

• Plans + generates mobile screens via Daisy's /api/projects/:id/runs orchestrator
• Iterates on screens by sending natural-language runs, or pushes direct HTML edits through the batch screen endpoint
• Manages projects (create, list, update, delete)
• Tunes themes (via PATCH /api/projects/:id) so screens look consistent
• Reads rendered HTML straight off each screen and renders PNG/JPEG images via POST /api/screenshots

The agent reports the credit cost from each run's creditsCharged. Defaults to async runs (wait="none") so long generations don't block your terminal.

Install

Claude Code

mkdir -p ~/.claude/skills
git clone https://github.com/<your-org>/daisy-skill ~/.claude/skills/daisy

Claude Code auto-discovers it on next launch. Verify with /skills and look for design-mobile-apps-with-daisy.

Cursor / other agent runtimes

Most agent skill catalogs read a SKILL.md with YAML frontmatter from a folder. Point your runtime at this repo and follow its docs.

Prerequisites

1. A Daisy Pro or Max subscription (https://www.daisy.now/pricing). Free/Plus tiers can't call the API.

2. An API key from Settings → API keys in the daisy dashboard.

3. One environment variable:

   export DAISY_API_KEY="dsy_live_..."

   Add it to your shell profile so the agent picks it up automatically. The domain is always https://www.daisy.now — there is no base-URL override.

Things to ask your agent

After installing, try any of these:

• "Design onboarding for a meditation app: welcome, breathing-style picker, and notifications opt-in."
• "Add a profile screen to the Runr project. Avatar, stats grid, and an edit button."
• "On the settings screen, change the primary button to red and make the destructive action stand out more."
• "Set a darker theme for this project and regenerate the home feed."
• "What did I create on Daisy this week? Render each screen to a PNG."
• "Cancel the run that's been going for 8 minutes."

The agent will pick the right endpoints, batch work into runs when sensible, save the rendered HTML or screenshots locally for you, and report the credits each run charged.

How it works

The skill is built around progressive disclosure — the lean part loads first, the rest only when needed:

skills/
├── SKILL.md                  # The one path: run → poll → HTML → image, plus guardrails
└── reference/
    ├── api.md                # Endpoint table, scopes, errors, rate limits, idempotency, credits
    └── direct-ops.md         # Advanced: single-screen create, batch edits, theme, embedding

At startup your agent loads only the name + description from the frontmatter. When a prompt matches, it reads SKILL.md — the opinionated happy path it follows for ~90% of requests. It opens a reference/ file only when it hits an error or needs the advanced path, so idle token cost stays near zero.

Together these teach your agent:

• The single recommended workflow: the runs orchestrator, async, polled
• How to authenticate, handle errors, retry safely, and respect rate limits
• Credit costs per operation so the agent can be honest about budget
• The full request/response shape for every endpoint, with copy-pasteable curl
• When to drop down to direct screen ops for precise control

Security notes

• API keys are scoped (full_access / read_only / custom). The agent will respect whatever scopes you grant — pick read_only (projects:read, screens:read, runs:read) if you want it to inspect but never spend credits.
• Screens return their rendered HTML directly, and screenshots are fetched with your key — there are no public preview URLs to leak.
• The skill never asks the agent to store your API key — it always reads from DAISY_API_KEY. If your agent tries to print or persist the key, that's a bug in the agent, not the skill.

Limits

• 1 active run per project (409 on concurrent starts)
• Per-key rate limits (429 with Retry-After; see the X-RateLimit-* response headers)
• Batch screen update/delete: 1–500 items per request
• Only screens with status done can be rendered to an image (409 otherwise)

Links

• Daisy: https://www.daisy.now
• Pricing: https://www.daisy.now/pricing
• Dashboard: https://www.daisy.now/dashboard
• Issues with the skill: open one in this repo
• Issues with the API itself: support at daisy.now

License

MIT. Use it, fork it, embed it in your own agent runtime.