Skip to main content
AI-agent skills and knowledge for building on the SODAX cross-chain DeFi platform. Drop this into your repo and your AI coding agent writes v2-correct @sodax/* SDK code on the first try. Full setup (skills CLI, npm, monorepo/local install, wiring agents to AGENTS.md): docs/ai-integration-guide.md.

​
Install

Using the skills CLI from Vercel Labs — the open agent-skills ecosystem CLI (supports Claude Code, Cursor, Codex, GitHub Copilot, and 50+ other agents):
The five installable skills listed in .claude-plugin/plugin.json land in your repo — the sodax-build front-door / ideation skill plus the four mode-gated per-SDK-package skills (each with two knowledge subtrees: integration/ for new v2 code, migration-v1-to-v2/ for v1→v2 porting). The CLI does not copy the package-level AGENTS.md; agents auto-discover each skill from its SKILL.md frontmatter. Re-running the command picks up the latest content.
npm fallback (web chats or when you prefer a devDependency): pnpm add -D @sodax/skills, then point your agent at node_modules/@sodax/skills/AGENTS.md. See the integration guide.

​
What you get

Skills are short and action-oriented (workflow + anti-patterns + links). Knowledge is the lookup material. Don’t read knowledge files top-to-bottom — the skill tells the agent which file is relevant for the current task.

​
Which skill applies?

After install, your agent picks based on what you’re building. Quick guide: AGENTS.md says the same thing in router form — your agent reads it first and picks.

​
Why this exists

LLM training data drifts: snippets from chat often use stale method names, reshaped types, or outdated error codes. Public docs help humans, not agents — an agent only reads what’s in its context window. This package ships the right material in agent-native form so the agent reads it before generating code. The content is version-locked to the SDK — upgrade @sodax/skills, the docs upgrade with it.

​
Feedback

If your agent generates wrong code despite reading the docs, that’s a doc bug — please open an issue on the SODAX SDKs repo with the prompt and the incorrect output. The per-skill knowledge/ subtrees are structurally CI-guarded (frontmatter, link resolution); prose claims benefit from real-world feedback.