AIToday
Large Language ModelsAI Coding AssistantsOpen-Source AIZenn AI/MLPublished: Nov 17, 2026, 16:00 JST

kensaurus/skills ships a 完了ゲート to block agents' false "done" claims

Send on LINE
kensaurus/skills ships a 完了ゲート to block agents' false "done" claims

3 Key Points

  1. What happened

    kensaurus published kensaurus/skills, 162 Markdown skill files under an MIT license, built around a completion gate that reads a state file and sends agents back if items are unchecked.

  2. Why it matters

    Because agents often declare done when only compiling, the gate checks completion from outside and a separate judge agent weighs the original request, the file, the diff and fresh test results.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The release sits on top of Agent Skills, a public specification that lets the same SKILL.md file be read by Claude Code, Cursor, Codex CLI, Gemini CLI and GitHub Copilot, according to kensaurus. The stated problem is not a lack of agent capability but of follow-through: agents call work done once a diff compiles, skip tests, or fill spec gaps on their own.

The gate tackles this with three parts: a checklist kept in a file rather than in the agent's context, a Stop hook that pushes the agent back when items remain, and a judge sub-agent that starts with fresh context so it has to find evidence rather than trust memory. In the v2.8.0 release, the judge returned CONTINUE with six fixes even though CI passed; one fix concerned an effort setting that bled into other skills run in the same turn.

kensaurus places the collection honestly against more popular alternatives, comparing GitHub stars and noting the four comparison repositories are far larger in users. A study by Song & Wei reports a success-rate drop of up to 21% across a 202 library, which suggests that more skills can raise mix-ups rather than reduce them.

FAQ
How does the completion gate know work is unfinished?
The complete-everything skill first writes a checklist to .cursor/complete-everything-state.md. A Stop hook then reads the file and sends the agent back if any item is unchecked, while skips without a stated reason count as unfinished.
Do I have to install all 162 skills?
No. You can run the installer with the core profile, which covers about 30 skills for daily work and keeps the description list to under one fifth of the total.
Is there evidence the gate improves completion rates?
Not yet. kensaurus prepared 12 evaluation cases for whether the right skill gets called but has run only one, and says the data on how much the gate raises completion rates is still to come.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleRevolut: Codex cuts coding time 30–60%, GPT‑5.5 runs FinCrime checks