DooDooLamb News
Show HN: Poka-Yoke – mistake-proofing skills for Claude Code, with a benchmark
Brief published August 26, 2026 · Original source published August 25, 2026
Original reporting by rainmanjam at github.com.
Automated brief. Verify important details at the original source.
What happened
A GitHub project applies Shigeo Shingo's poka-yoke manufacturing concept to AI-assisted coding via a Claude Code plugin. The tool audits code for error-prone patterns, designs APIs to resist misuse, installs guardrails, and converts past incidents into preventive devices. A benchmark accompanying the release measures how often models proactively name what a proposed fix forecloses: without the plugin, Claude does this 42% of the time; with the skills enabled, that rate rises to 81%.
Why it matters
The benchmark surfaces a concrete gap in how coding agents communicate tradeoffs. An agent that closes a design without flagging what it makes impossible leaves developers with incomplete information. The jump from 42% to 81% unprompted disclosure suggests structured prompting or skill scaffolding can meaningfully shift model behavior on a measurable safety-relevant dimension.
What to watch
Whether the 81% figure holds across model versions and codebases beyond the benchmark set, and whether other agent frameworks adopt similar tradeoff-disclosure evaluations as a standard quality signal.