Set Up Your AI Tools
Install CCC's Agent Skills into Cursor, Claude Code, GitHub Copilot, Windsurf, and 60+ other AI tools with one command.
Your AI assistant only writes correct CCC code once it's actually loading CCC's rules. CCC ships these as a set of Agent Skills — one hub skill, ckb-ccc-fundamentals (package selection, Cell model, address/amount handling, hallucination guard), plus a spoke skill per task area (ckb-ccc-signer-setup, ckb-ccc-transactions, ckb-ccc-udt, ckb-ccc-spore, ckb-ccc-playground, ckb-ccc-examples-finder), all committed under skills/ in the CCC repo.
Install them with skills, the open Agent Skills CLI — it auto-discovers the skills/ directory and writes each skill to the right path for whichever tool(s) you target:
npx skills add ckb-devrel/ccc --all--all installs every skill non-interactively — the one-line path most projects want. Leave it off and the CLI opens an interactive picker for which skills to install (in addition to the tool picker), which is useful if you only want a subset:
# Only Cursor and Claude Code, every skill
npx skills add ckb-devrel/ccc --all -a cursor -a claude-code
# Every supported tool, only the fundamentals + UDT skills
npx skills add ckb-devrel/ccc --skill 'ckb-ccc-fundamentals' --skill 'ckb-ccc-udt'
# Install globally (all your projects) instead of this project only
npx skills add ckb-devrel/ccc --all -gskills supports Cursor, Claude Code, GitHub Copilot, Windsurf, Codex, and 60+ other agents — each skill's description frontmatter drives when the tool loads it, so no extra rule-file wiring is needed on top of the install.
Do this once per project (or once globally with -g). It takes about two minutes and is the single highest-leverage thing you can do to improve AI-generated CCC code. After setup, always run the check on Verify & Troubleshoot — a skill that silently failed to load is worse than none.
For tools without a file-based rules/skills mechanism (e.g. plain ChatGPT), npx skills add has nothing to write to. Paste this into the system prompt (or the first message) instead:
Before writing any CKB/CCC code, fetch https://docs.ckbccc.com/skill.md to see the list of CCC agent skills.
Always fetch and follow the ckb-ccc-fundamentals skill first,
then fetch the spoke skill matching the task (signer setup, transactions, UDT, Spore,
playground, or examples) before generating code.Keep skills up to date
An installed skill is a point-in-time copy — CCC's skills keep improving, and nothing pushes those improvements to a project that already ran npx skills add. skills tracks what you installed in a skills-lock.json, so you can check and pull updates on demand:
# Pull the latest version of every installed skill
npx skills update
# Update just one
npx skills update ckb-ccc-udtnpx skills update diffs content hashes, not version numbers, and has a known gap where it can report "up to date" even when the source changed (vercel-labs/skills#484). If a canary question on Verify & Troubleshoot starts failing right after a CCC skills update was announced, don't trust check alone — remove and re-add the skill (npx skills remove ckb-ccc-fundamentals && npx skills add ckb-devrel/ccc --skill 'ckb-ccc-fundamentals').
There's no notification when a skill changes — check periodically (e.g. before starting significant new work) rather than assuming you'll hear about it.
Tell your AI where to find answers
The skills cover the common cases, but for anything beyond them your assistant should fetch from the docs rather than guess. These are the machine-readable entry points to point it at — the same URLs work whether you paste them into a prompt or bake them into a rule:
| Resource | URL | Use it for |
|---|---|---|
| Skills index | /skill.md | The default: lists every skill (hub + spoke), what each covers, and its raw SKILL.md URL. Load ckb-ccc-fundamentals first. |
| Docs index | /llms.txt | "Which page covers topic X?" A short, linked index of every docs page — fetch first to navigate. |
| Single page as Markdown | any docs URL + .md | A focused how-to on one topic, e.g. https://docs.ckbccc.com/en/docs/guides/udt-tokens.md. Clean Markdown, no page chrome. |
| Full docs dump | /llms-full.txt | A broad question spanning several topics, when you want everything in one fetch. |
| API reference | api.ckbccc.com | Exact method signatures, parameter types, and enum values for any @ckb-ccc/* export. |
Any docs page has a Markdown twin: append .md to its URL (or send an Accept: text/markdown header). That's what you want an AI to read — it's the article text without navigation or component markup, so it spends context on content, not layout.
Verify it worked
Ask the assistant a CKB-specific question it would otherwise get wrong:
"In CCC, does
signer.getBalance()return anumberor abigint?"
A correctly configured assistant answers bigint (in Shannon) — this exact gotcha is called out in the ckb-ccc-fundamentals skill's Hallucination Guard. If it answers number, the skill/rule isn't loading. See Verify & Troubleshoot for four more of these canary questions and a full decision tree for fixing a broken setup.
Next step
Configuring the tool is necessary but not sufficient — see Prompting Best Practices for how to phrase requests so the assistant actually uses what it now has access to.