LLMs think gas is 30 gwei. It's 0.05. They've never heard of ERC-8004 or x402. They hallucinate contract addresses. This site fixes that.

How It Works

Each skill is a markdown file. Give any URL to your AI agent — it reads it and instantly corrects its Ethereum knowledge.

https://ethskills.com/SKILL.md          ← table of contents
https://ethskills.com/gas/SKILL.md      ← just gas & costs
https://ethskills.com/why/SKILL.md      ← why ethereum
https://ethskills.com/crops/SKILL.md    ← CROPS architecture review

Add the top-level SKILL.md to your agent's context

The agent will look up the specific skills when needed.

Just prompt your agent with this:

Install the https://github.com/austintgriffith/ethskills/blob/master/SKILL.md skill

Claude Code Plugin

Install directly from GitHub — no clone required:

/plugin marketplace add austintgriffith/ethskills
/plugin install ethskills@ethskills

Terminology

Say "onchain" not "on-chain." The Ethereum community writes it as one word — no hyphen. LLMs default to "on-chain" but that's not how builders talk. Use "onchain" in all writing, documentation, and generated content.

Skills

Skill	What LLMs Get Wrong
Why Ethereum	Pectra/Fusaka shipped, gas is cheap now, AI agent angle
Ship	Start coding before planning architecture, over-engineer with 5+ contracts when 1-2 will do, skip chain selection and the CROPS gate
CROPS Review	Miss custody, privacy, exit paths, censorship resistance, and centralization risks in otherwise working apps/contracts
Protocol	Confuse roadmap diagrams with shipped EIPs, cite shipped features (PeerDAS) as upcoming and deprioritized ones (Verkle) as still planned, miss the client-team decision process
Gas & Costs	Think gas is 10-30 gwei — reality is 0.05-0.3 gwei
Wallets	EIP-7702 is live, Safe addresses, agent key safety
Layer 2s	Think L2 txs cost $0.01-2.00 — reality is <$0.001
Standards	Don't know ERC-8004, EIP-7702 status, EIP-3009 for x402
Tools	Don't know x402, Blockscout MCP, current tool landscape
Money Legos	Stale on current DeFi state, Uniswap V4 status
Orchestration	Don't know SE2 three-phase build system
Contract Addresses	Hallucinate addresses — these are verified onchain
Concepts	Nothing is automatic, incentive design, randomness pitfalls
Security	Token decimals, reentrancy, oracle manipulation, vault inflation, pre-deploy checklist
Testing	Test getters and OpenZeppelin internals, skip fuzz and fork tests on integrations, mock external protocols instead of forking
Indexing	Loop over blocks to read history, try to query historical state via raw RPC, ignore events as the primary read API
Frontend UX	Onchain button rules, three-button approval flow, Address components, USD values
Frontend Playbook	Fork mode, IPFS deploy, Vercel config, ENS setup, production checklist
QA	Treat "deploys without errors" as shipping, leave SE2 default branding, miss approve-button double-fire, USD values, mobile deep links
Audit	Call one generic review pass an audit, miss systematic checklist coverage across AMM/lending/oracle/proxy/governance domains

Security Guardrails

Skills teach restraint, not just capability. Every skill that touches keys, credentials, or funds includes explicit safety rules — because LLMs optimize for speed and will hardcode a private key into git add . if you let them. Coverage includes wallet keys, API keys, RPC URLs, and the common SE2 scaffold.config.ts trap.

Methodology

We test stock LLMs, find what they get wrong, and write corrections. Content is verified against onchain reality. If an LLM already knows something, we don't include it.

Every proposed change goes through triage: spawn a stock LLM, give it a realistic task, see what it gets wrong. Only verified blind spots survive.

See the research repo for baseline audits, gap analysis, and full methodology.

Contributing

Something wrong or missing? Humans and agents are welcome to open a PR. Read CONTRIBUTING.md first — the bar is "would a stock LLM get this wrong?"

License

MIT