Nine skills we wish our agents shipped with

Every team that adopts coding agents hits the same three problems within weeks, and none of them is model quality. The agent says "done" and nobody can tell what that was checked against. A session ends and the next one starts from a vague summary. A significant decision gets made by one model in one context window with nobody challenging it.

We build software with agents every day at Agentic Engineering, and we kept paying for those three. So we turned our fixes into nine small skills and published them under the MIT license: github.com/luci-efe/skills.

• /doctor lists missing tools with the install command for each.

• /compass routes a task to the right skill and checks tools first.

• /cartographer maps one flow to verified symbols and tests.

• /crucible reads the code first and asks only the open questions.

• /council sends one artifact to several model families for two rounds of review.

• /reproduce requires a failing test before the fix.

• /proof records checks against the exact code state, and reports evidence stale if a source file changes afterwards.

• /red-team triages dependency advisories into confirmed, suspected, not affected, not found within stated coverage or unknown.

• /relay hands work between sessions as a checkpoint that can't be edited quietly.

The suite doesn't make agents correct. It makes their claims checkable. When a tool is missing, a skill stops and prints the install command; nothing silently degrades.

We tested all of it on eight real repositories, not fixtures. The public runs are on our own ultimate-harness repo, and they surfaced bugs we then fixed: the suite's tests went from 99 to 127.

What we'd tell a client before they install it:

• The hard /proof gate is wired for Claude Code and Codex (watched live on Codex) and, from its docs, Gemini CLI (not watched live). Elsewhere the scripts run as manual checks.

• It's a local gate. We call a result "done" when it passes and "verified" only after CI re-executes the checks and an independent reviewer looks.

• /council has run with two model families so far, not three.

npx skills@latest add luci-efe/skills

Run /doctor afterwards. If your team is adopting agents and wants verification, handoff and review set up in your own workflow, talk to us at agenticengineering.agency. The skills are free and will stay free. We'd rather have them used and criticized than admired.