Nine skills I wish my agents shipped with

My agent told me it was done. I asked what "done" meant. It told me again, with more confidence.

That's the whole origin story. Agents say "all tests pass" and nobody can tell which tests, on which commit, or whether the code moved afterwards. Sessions end and the next one starts from a vague summary. A big decision gets made by one model in one context window, and nobody argues with it. So I wrote nine small skills that fix those three things, put an MIT license on them and published the lot: github.com/luci-efe/skills.

Here's the cast:

/doctor lists missing tools and prints the install command for each.
/compass routes a task to the right skill, after checking the tools exist.
/cartographer maps one flow to real symbols and tests.
/crucible reads the code first and asks only what the code can't answer.
/council sends one artifact to seats from different model families, two rounds, a chair writes the synthesis.
/reproduce wants a failing test before any fix.
/proof records your checks against the exact code state. Touch a file afterwards and the evidence goes stale.
/red-team scans dependencies and refuses to call a scanner count a finding.
/relay hands work to another session as a checkpoint you can't quietly edit.

Is it magic? No. It makes an agent's claims checkable, which is a lot less glamorous and a lot more useful. If a tool is missing, the skill stops and tells you what to install. Nothing silently degrades, because silent degradation is how you end up with a demo that works.

I didn't test them on fixtures I wrote to flatter myself. I ran every skill on real repositories, eight in total, and the public write-ups are on one of ours, ultimate-harness, a TypeScript and Python project with close to 900 tests. It found bugs in my code. The suite's own tests went from 99 to 127 along the way.

The limits, up front. The one hard gate (the /proof stop hook) is wired for Claude Code and Codex, and I watched Codex block a stale "done" live. Gemini CLI is wired from its docs and not watched live. In Cursor, Copilot and OpenCode the same scripts run as manual checks. The gate is local, so hand-written evidence can still pass it; I call that "done", and "verified" needs CI re-running the checks plus a reviewer who isn't the agent. And /council has only run with two model families, not three. Gemini ran out of quota. Cry about it.

npx skills@latest add luci-efe/skills

Then run /doctor in your repo. Want fewer skills? -s proof -s relay works.

Bug reports, war stories and gentle abuse: open an issue, or find me in r/AgenticEngineeringCo.

Fernando Ramos, co-CEO of Agentic Engineering. More at ferramos.com.