SuperAPI AI Directory
Partner中文AI newsFavorites
Skillsdesign-agent-customize

design-agent-customize

Design & creation

View repository

Updated 10-28 21:24

About this skill

Customize Design Agent

Every kind of knowledge has one home. Find the owner before editing, so a fix lands once and every asset after it follows:

The change Its home
Brand facts: colours, fonts, logos, voice, graphic language workspace/brand/ (brand-onboarding)
A deliverable or workflow the team asks for often a runtime skill, skills/<name>/SKILL.md
Doctrine for every visual: hierarchy, type, colour, charts skills/design-wiki/references/
Identity, delivery and approval rules, how it talks instructions.md (a budget test fails past about 8k tokens)
A rule a machine can count a check in workspace/tools/
A tool or data source an MCP server in config/mcp.toml
Which model plays which role, effort, budgets .env settings; Models

Add a runtime skill

Start with the graphic the team asks for every week.

  1. Read two skills close to it (skills/stills/, skills/thumbnails/) and skills/design-wiki/SKILL.md.
  2. Create skills/<name>/SKILL.md as a real folder one level deep (MDA does not follow symlinks there). Frontmatter: name is the folder; description says when to use it, in the words people use when they ask. Body: what good looks like, sizes, steps, the checks to run. Put depth in references/ and keep the body small. Text only: binary files (fonts, images) go in skill-binaries/<name>/, because Context Hub takes only UTF-8 text; the stager puts them back beside the skill in the sandbox.
  3. No brand facts, colours or libraries from a CDN in the skill: the brand pack supplies them.
  4. uv run pytest -q tests/unit/test_taste_skills.py tests/unit/test_media_skills.py checks frontmatter, size, links, and that every tool, script and check it names exists.
  5. Try it with the request people actually send: uv run design-agent ask "…". Look at the files.
  6. Add an eval task if it matters (design-agent-evals), then redeploy (design-agent-deploy).

Vendored skills (HyperFrames and LangSmith, listed in skills-lock.json) are never edited: override them in a sibling skill of ours, or send the change upstream.

Instructions, doctrine and checks

  • Phrase rules as defaults with a reason, or as checks with evidence. A rule about the model is temporary; retest it when the model changes and delete what the new model outgrows.
  • A rule a machine can count belongs in code, as a check, not in prose.
  • After editing, run uv run pytest -q, then compare before and after with the evals.

Tools

  • MCP servers: copy config/mcp.example.toml to config/mcp.toml and declare servers. Locally, header values come from the environment (headers_env); on MDA a server that needs credentials names a workspace connection (connection), owned by the deployment or each person (owner). The example file has presets for Notion, Linear, GitHub, Figma and others (integrations). MDA serves HTTP and SSE only. Paper is configured with PAPER_*, not here (Paper).
  • Media keys: OPENAI_API_KEY (images), ELEVENLABS_API_KEY (voice, music, sound effects), GROQ_API_KEY (transcription).
  • A new host tool is code in src/design_agent/tools/: keys stay on the host and results drop into the sandbox as files. Follow the change discipline in AGENTS.md.

Models and cost

uv run design-agent config set DESIGN_AGENT_MODEL=provider:model, then uv run design-agent test model. DESIGN_AGENT_JOB_BUDGET_USD makes the agent ask before passing a budget; DESIGN_AGENT_HELPER_EFFORT_CAP caps helpers' effort. More: Customization.