Cds Docs Agent Skill

CDS Documentation Skill

Goal: Pull CDS documentation into the session so your next edits, recommendations, and

API usage are grounded in current docs—not to produce a documentation handout for the user.

Treat fetched pages as source material you have read: internalize props, patterns, imports, and

setup constraints, then act (code, refactors, answers) from that understanding. Quote or

summarize for the user only when it helps. Keep the actual doc page contents to yourself

and only share their contents directly to the user if they ask.

Whether you use the CDS MCP server or fetch the pages yourself with curl is an implementation detail.

The use does not need to know your methodology unless they indicated they are debugging this skill.

Ambiguous product requests (e.g. “turn this into tabs,” “make it a card”) after you’ve read the

index: if several CDS components could be relevant (e.g. underline Tabs vs pill SegmentedTabs vs

TabNavigation), it is acceptable and often better to ask one short, concrete question (e.g.

which visual pattern matches their spec) while briefly naming the options and how they differ per

the docs. Do not pick one arbitrarily just to avoid asking.

1. Choose platform

Decide whether the work targets web (React, @coinbase/cds-web) or mobile (React Native,

@coinbase/cds-mobile). If unclear, infer from the repo (e.g. package.json dependencies, apps/*

layout) or ask a single clarifying question.

Pass web or mobile to list-cds-routes and when building curl URLs.

2. Tool order (try in sequence)

• CDS MCP server (packaged mcp-docs in @coinbase/cds-mcp-server / packages/mcp-server):
• list-cds-routes with platform: "web" | "mobile".
• get-cds-doc with route: path including the platform, ending in .txt (e.g.

web/getting-started/installation.txt). Paths must come from the index.

• curl if MCP is not configured, fails, or returns errors:
• Index: https://cds.coinbase.com/llms/web/routes.txt or

https://cds.coinbase.com/llms/mobile/routes.txt

• Page: https://cds.coinbase.com/llms/<route> where <route> matches the index (e.g.

web/getting-started/theming.txt).

Use curl -fsSL '<url>'. Prefer targeted fetches (the pages you need for this task) over

loading everything; the index is for routing, not for echoing back wholesale.

3. Workflow whenever this skill applies

• Resolve platform (web or mobile).
• Load the index once per platform focus (list-cds-routes or routes.txt) and use it as the

canonical map of .txt paths.

• Fetch additional pages only as needed for the work ahead; after each fetch, apply what

you learned (imports, props, patterns)—do not stop at regurgitating the file.

4. Practical notes

• Routes are case-sensitive; doc content is .txt from MCP or HTTP.
• Web and mobile docs differ; confirm the component or topic exists for that platform’s index.
• Human-readable URLs inside the .txt files are optional to share with the user; your priority is

correct implementation, not reproducing the doc site in the reply.

5. What “done” looks like

• Platform is correct or explicitly chosen.
• Index informed which pages you pulled.
• User-facing output is implementation-focused (components, imports, structure); retrieval steps

stay implicit unless the user needs them.

• If multiple components fit, you either ground the choice in doc differences or **ask one

clarifying question** when the product spec is under-specified.

• Your actions and explanations reflect fetched CDS details (APIs, setup, tokens)—especially

where memory would be risky—not a dump of documentation for its own sake.