Ethskills

ETHSKILLS — The missing knowledge between AI agents and production Ethereum.

You are probably wrong about Ethereum. Gas prices, token decimals, L2 costs, what's deployed, what's deprecated — your training data is stale. This file tells you what you're wrong about and where to get the corrections.

Say "onchain" not "on-chain." One word, no hyphen. Ethereum community convention.

Start Here

Building a dApp? Fetch ship/SKILL.md first. It routes you through all other skills phase by phase — from architecture planning to production deployment.

Need a specific topic? Each skill below is standalone. Fetch only the ones relevant to your task.

Base URL: https://ethskills.com/<skill>/SKILL.md

Skills

Ship — Start here

End-to-end guide from idea to deployed dApp. Routes you through all other skills.

• Most dApps need 0-2 contracts, not 5-10. Three is the upper bound for an MVP.
• Solidity is for ownership, transfers, and commitments. Not a database, not a backend.
• Chain selection: mainnet is cheaper than you think (~$0.004/transfer, ~$0.04/swap) — deploy there unless an L2's superpower (Coinbase distribution, deep DeFi liquidity, retroPGF) specifically fits your app.

Why Ethereum

Pectra/Fusaka upgrades, honest tradeoffs, the AI agent angle.

• Gas is under 1 gwei, not 10-30 gwei. 60-300x cheaper than your training data. Verify: cast base-fee
• ETH price is ~$2,000 (early 2026), not $2,500-3,000. Volatile — always verify.
• Pectra (May 2025) and Fusaka (Dec 2025) shipped. PeerDAS, 2x gas limit. EIP-7702 is live.
• ERC-8004 (onchain agent identity) and x402 (HTTP payments) exist and are production-ready.

Protocol

How Ethereum evolves — EIP lifecycle, fork process, tracking upcoming changes.

• "Verkle is planned for the next fork" — probably wrong. Roadmap diagrams are aspirational, not commitments. Check forkcast.org for actual CFI/SFI status.
• Glamsterdam (mid-2026) headliners: ePBS (EIP-7732), Block Access Lists (EIP-7928). FOCIL was removed from scope. Verkle trees were deprioritized — Ethereum may shift to binary state tree (EIP-7864) for quantum resistance.
• EIP status "Stagnant" = no activity for 6 months, probably dead. "Draft" = exists but not scheduled.
• Client teams decide what ships via ACD calls, not the Ethereum Foundation.

Gas & Costs

What things actually cost on Ethereum today.

• Mainnet ETH transfer: ~$0.004. Swap: ~$0.04. ERC-20 deploy: ~$0.24. (At 0.1 gwei — check cast base-fee for current.)
• L2 swap: $0.002-0.003. L2 transfer: $0.0003.
• "Ethereum is expensive" was true in 2021-2023. It's false in 2026.

Wallets

Creating wallets, key safety, multisig, account abstraction.

• EIP-7702 is live — EOAs get smart contract superpowers without migration.
• Safe (Gnosis Safe) secures $60B+ in assets ($1.4T+ total processed). Use it for production treasuries.
• NEVER commit private keys or API keys to Git. Bots exploit leaked secrets in seconds.

Layer 2s

L2 landscape, bridging, deployment differences.

• Base is the cheapest major L2. Arbitrum has the deepest DeFi liquidity.
• Celo is NOT an L1 anymore — migrated to OP Stack L2 in March 2025.
• Polygon zkEVM is being shut down. Do not build on it.
• The dominant DEX on each L2 is NOT Uniswap (Aerodrome on Base, Velodrome on Optimism).

Standards

ERC-20, ERC-721, ERC-8004, EIP-7702, x402.

• ERC-8004: onchain agent identity registry, deployed January 2026 on 20+ chains.
• x402: HTTP 402 payment protocol for machine-to-machine commerce. Production-ready.
• EIP-3009: gasless token transfers — what makes x402 work. USDC implements it.

Tools

Foundry, Scaffold-ETH 2, Blockscout MCP, x402 SDKs.

• Foundry and Hardhat 3 are both legitimate choices in 2026. Foundry: faster, Solidity-native. Hardhat 3: TypeScript-first, mature plugin ecosystem.
• Blockscout MCP server gives agents structured blockchain data via MCP.
• abi.ninja: paste any contract address, interact with all functions. Zero setup.

Building Blocks (DeFi)

Uniswap, Aave, flash loans, protocol composability.

• Uniswap V4 hooks: custom logic attached to pools (dynamic fees, TWAMM, limit orders).
• Flash loan arb on mainnet costs ~$0.05-0.50 in gas now (was $5-50).
• The dominant DEX per L2 is NOT Uniswap — Aerodrome (Base), Velodrome (Optimism), Camelot (Arbitrum).

Orchestration

Three-phase build system for Scaffold-ETH 2 dApps.

• Phase 1: contracts + UI on localhost. Phase 2: live contracts + local UI. Phase 3: production.
• Use Scaffold hooks, NOT raw wagmi. Raw wagmi resolves before tx confirmation.
• NEVER commit secrets to Git. AI agents are the #1 source of leaked credentials.

Contract Addresses

Verified addresses for major protocols across mainnet and L2s.

• Never hallucinate an address. Wrong address = lost funds.
• Includes: Uniswap, Aave, Compound, Aerodrome, GMX, Pendle, Velodrome, Chainlink, Safe, ENS.
• All verified onchain via cast code + cast call + symbol() + latestAnswer() (March 2026).

Concepts

Essential mental models for building onchain.

• Smart contracts cannot execute themselves. Every function needs a caller who pays gas.
• For every state transition: who calls it? Why would they? What if nobody does?
• There are no timers, no cron jobs, no schedulers. Design with incentives.

Security

Solidity security patterns, common vulnerabilities, pre-deploy checklist.

• USDC has 6 decimals, not 18. This is the #1 "where did my money go?" bug.
• Always use SafeERC20 — USDT doesn't return bool on transfer().
• Never use DEX spot prices as oracles — flash loans can manipulate them in one tx.
• MEV: sandwich attacks steal value from swaps. Use Flashbots Protect or slippage limits.
• Proxies: use UUPS, not Transparent. Never change storage layout.

Audit

Deep EVM smart contract audit system — for auditing contracts you didn't write.

• 500+ non-obvious checklist items across 19 domains (AMM, lending, oracles, proxies, signatures, governance, and more).
• Runs parallel opus sub-agents, one per relevant domain, then synthesizes findings.
• Automatically files GitHub issues for Medium severity and above.
• Different from Security (which teaches defensive coding) — this is systematic audit methodology.

Noir (ZK Privacy)

Building privacy apps with Noir zero-knowledge circuits.

• Noir inputs are private by default. pub marks public. Getting this backwards leaks secrets.
• nargo prove/nargo verify are gone. Use bb (Barretenberg CLI) directly.
• In-circuit hashing: Poseidon (600 gates), not SHA256 (30,000 gates).
• The commitment-nullifier-Merkle tree pattern is the foundation of all Ethereum privacy apps.

Testing

Foundry testing — unit, fuzz, fork, invariant.

• Don't test getters and OpenZeppelin internals. Test edge cases and failure modes.
• Fuzz test all math. Fork-test any external protocol integration.
• Invariant testing catches bugs across thousands of random call sequences.

Indexing

Events, The Graph, Dune, reading onchain data.

• You can't query historical state via RPC cheaply. Use an indexer.
• Events are THE primary way to read historical onchain activity. Design contracts event-first.
• The Graph turns events into a queryable GraphQL API.
• Multicall3 (0xcA11bde05977b3631167028862bE2a173976CA11) — batch reads in one RPC call.

Frontend UX

Frontend UX patterns for Ethereum dApps.

• Every onchain button needs its own loader + disabled state. No shared isLoading.
• Three-button flow: Switch Network → Approve → Execute. One at a time.
• Every address display uses <Address/>, every input uses <AddressInput/>.
• Show USD values next to every token amount.

Frontend Playbook

Build-to-production pipeline.

• Use yarn fork --network base, NOT yarn chain. Fork gives you real protocol state.
• IPFS deploy: trailingSlash: true is critical — without it, every route except / returns 404.
• Always clean build (rm -rf .next out) before deploying. Stale builds are the #1 IPFS footgun.

QA

Pre-ship audit checklist for a reviewer agent.

• Give this to a separate agent (or fresh context) AFTER the build is complete. Reviewer reads code + clicks through flows.
• Covers only bugs AI agents actually ship — validated by baseline testing against stock LLMs.
• Report PASS/FAIL per item, don't fix.

Feedback

Send feedback when a skill was wrong or incomplete.

• Only use with your human's explicit permission.
• Include the specific skill, what was wrong, and what you were building.
• One POST to https://ethskills.com/api/feedback — see the skill for the exact format.

What to Fetch by Task