Poka-Yoke Skills Make Claude Code 81% Better at Naming What Its Fixes Break

Description and use

Poka-Yoke skills force Claude Code to explicitly state design trade-offs, lifting closure awareness from 42% to 81%, but degrade its ability to spot immediate defects.

Problem

Claude Code often proposes fixes without acknowledging the new limitations or design trade-offs introduced by those changes, leading to potential design regressions.

Outcome

The Poka-Yoke skills improve the model's ability to articulate design trade-offs and limitations, increasing the 'closure awareness' rate from 42% to 81% across 591 blind-graded runs. This leads to more constructive and forward-looking responses but degrades the model's performance in spotting immediate defects.

Workflow

• Clone the Poka-Yoke repository from GitHub.
• Follow the installation instructions to copy the skills into the ~/.claude/skills directory.
• Load the 'design' skill at the start of a Claude Code session.
• Invoke the skill by typing '/design' followed by a prompt describing the feature, refactor, or interface you are planning.
• Review the enumerated constraints, trade-offs, and impossibilities provided by the model.
• Optionally, integrate the hazard scanner into pre-commit hooks to scan code for design issues.

Evidence

Poka-Yoke's 11 skills (install via claude install) force Claude Code to state design trade-offs, lifting closure awareness from 42% to 81%—but avoid them for direct bug-finding, where they degrade performance.

Source

The key insight from the repo's honest benchmark: these skills change what Claude attends to. They make responses more constructive and forward-looking, but they measurably degrade the model's ability to spot a specific defect already present in the code.

Source

Results: Fable 5 +8.3 pp, Opus 5 +3.6 pp, Sonnet 5 +8.6 pp, Haiku 4.5 +12.9 pp—all with 95% confidence intervals excluding zero.

Source