Skip to main content
A skill is a package of instructions that teaches your AI agent how to use the Taste Engine well. Where the MCP server gives your agent the tools, a skill gives it the judgment. Two skills ship today: brand-search finds real design references before the agent commits to a visual direction, and brand-adherence builds a page inside an existing brand’s identity, then verifies the result instead of trusting its own eyes. Skills follow the open SKILL.md standard, so the same skill works across Claude Code, Codex, Cursor, OpenCode, Gemini CLI, and other agents that support it.

​
Install

Install with the skills CLI. No global install is needed: npx runs it on demand.
A few variations:
You can also install manually by copying a skill folder into your agent’s skills directory:

​
Connect the MCP server first

npx skills add installs the instructions, not the tools. The skills drive the Taste Engine MCP tools for search, extraction, and verification, so they only work once the MCP server is connected. For example, with Claude Code:
If your agent does not see the tools, the MCP server is not connected. See MCP server for setup in Codex, Cursor, and other clients.

​
What the skills do

The brand-search skill triggers whenever your agent needs website design references: you describe a style or mood (“dreamy editorial skincare, calm and warm”), name reference sites (“layout of linear.app, colors of ramp.com”), ask for brands that look like an existing site (“competitors of acme.com”), or ask for a website with no visual direction at all (“build a website for my bakery”). Instead of inventing a look, the agent grounds its design decisions in real websites from the curated corpus, in four moves:
  • Pick the right search tool. A description becomes a search_brands query. A brand you already have becomes a search_similar_brands call on its extraction: the agent reuses an existing submission_id, or extracts the site first when it only has a URL. Sites you name are extracted directly, since they are already the reference. An unqualified site (“sites like patagonia.com”) gets both lenses: its visual neighbours from search_similar_brands, and its cultural neighbourhood from a search_brands query anchored in what the site is about.
  • Search what the prompt says, not what the agent assumes. The query is built from the words you wrote: industry, page type, audience, named references, and any style, layout, or palette terms. Your style words travel verbatim (“vintage” stays “vintage”, even where the corpus tags that register “retro”), and the agent adds no adjectives you never used. When the prompt names no style, the query names none either, and the results propose the directions. Empty adjectives (“completely custom and unique”) are dropped, and the one real fact left is what gets searched.
  • Inspect every result, in order. Results come back ranked, and every card is mandatory evidence, including the discovery results that widen the set. For each one the agent reads the metadata and description and views the screenshot, in the returned order, before choosing references. It never re-ranks on a hunch: claims about motion or interaction get checked on the live site, and a result is set aside only with a stated mismatch against your prompt.
  • Extract only when it needs to. A card’s metadata, description, and screenshot are usually enough to judge a reference, and often enough to design from. The agent runs extract_brand when it will build directly from a source and needs its real values, or when it needs a submission_id to feed search_similar_brands. It polls to completion, reads the result, and replaces any failed extraction with the next ranked result from the same query.
The skill asks for six results per query by default and raises top_k (up to 12) when you want a broad moodboard. Find inspiration walks the same search, pick, extract workflow over the HTTP API.

​
Ship on-brand pages with brand-adherence

The brand-adherence skill triggers when you name one reference site and ask for a page inside its identity: “a pricing page for stripe.com”, “a careers page in linear.app’s system”, “as if their design team built it”. It ships that page as if the brand’s own team shipped it, in three moves:
  • Pull the whole brand, exactly. The agent pulls the extraction a few sections at a time, saves each answer to disk, and copies values out of those files when it writes code: exact hex, exact type specs, exact CSS, with no transcription from memory. It also downloads the captured CSS, HTML, and screenshot, which carry the real font-loading rules, the real logo, and the honest picture of the brand’s density and mood.
  • Compose inside hard rules. Every color traces to a brand token, only the brand’s typefaces load, and every component is built from the brand’s own CSS. Layout, hierarchy, and copy are the agent’s latitude; the tokens are not.
  • Verify with the Verifier, twice. Once the page is on a URL the engine can reach, the agent runs a brand-adherence check, applies the fixes and recommendations, re-verifies, and closes with both scores side by side. You decide whether another round is worth it.
The Verifier extracts both sides from their URLs, so the finished page needs a URL the engine can reach. A page served only on localhost is invisible to it.

​
Available skills

The two chain naturally: brand-search finds and extracts the reference, and brand-adherence builds inside it. The catalog lives in the Taste-AI/skills repository and grows over time. Run npx skills add Taste-AI/skills --list to see the current set.

​
Next steps

MCP server

Connect the tools the skills rely on.

Find inspiration

The search, pick, extract workflow the brand-search skill runs.

Verification in an agent loop

The verify, fix, re-verify pattern the brand-adherence skill runs.