ETHSKILLS

Why it's ranked here

The format targets the specific failure of agents working on Ethereum: stale numbers and invented addresses. The README states a method, testing stock models on realistic tasks and keeping only corrections for verified blind spots, and the security sheet backs that with defensive code rather than warnings: decimal normalisation for six-decimal stablecoins, checks-effects-interactions with a reentrancy guard, oracle staleness checks, and a virtual-offset fix for vault inflation. It is MIT licensed and installs as a Claude Code plugin in two commands. The verdict is positive for Solidity work, with the caveat that its numeric content ages.

What's good

Guidance is concrete. The testing sheet says what not to test, such as library internals and getters, and shows fuzz, fork and invariant tests with bounded inputs. The wallet sheet lists ten key-safety rules for agents, including never moving funds without human confirmation and using a dedicated low-balance wallet, plus a two-of-three Safe pattern where a human can remove a compromised agent key. The planning sheet caps an MVP at three contracts and gives a per-function worksheet asking who calls it and what happens if nobody does. Every file that touches keys repeats the rule against committing secrets.

Tradeoffs

Much of the value is time-sensitive: gas prices, ETH price, fee tables and upgrade status are pinned to early 2026, and the gas sheet itself says to re-verify once its stamp is over 30 days old. The audit routine does not ship its checklists; it tells the agent to fetch them from a separate repository, spawn five to eight parallel sub-agents, and file GitHub issues automatically for Medium severity and above. The index and gas sheet ask the agent to send feedback to the project's own web endpoint and say to skip the permission ask if pre-authorized, which is an outbound network call worth gating.

How to use it well

Load the index into an agent that writes Solidity or builds a dApp and let it fetch topic sheets on demand; start new projects with the planning sheet, which routes through the rest phase by phase. Treat prices and cost tables as order-of-magnitude guides and check live values with the suggested command-line calls. Review the audit routine before running it, since it pulls outside checklists and opens issues. Deny or review the feedback request if your agent should not post to third-party endpoints. It does not replace a professional audit or live chain data.

Observed

License

MIT

Format

Markdown skill files with name/description frontmatter, one per topic

Install surface

Claude Code plugin marketplace, or a URL handed to any agent

Plugin manifest

.claude-plugin/plugin.json lists twenty skill directories

Audit checklists

Fetched from a separate external repository, not bundled

Outbound call

Feedback sheet instructs agents to POST to the project's feedback endpoint

Primary content

Markdown guidance for agents, not a library or CLI