Plugin name
poka-yoke
Short description
Mistake-proof your code, your pipeline, and your agents. Audits code for mistakes that are possible, designs APIs where misuse cannot be expressed, installs guardrails, and turns incidents into devices that prevent recurrence.
GitHub repository
rainmanjam/poka-yoke
Plugin path inside the repository
plugins/poka-yoke
Ref to review
v0.2.0
Commit SHA to review
18e2bc17d682ece3f13892e7415409a9da57d031
Version
0.2.0
License identifier
MIT
Author name
rainmanjam
Author URL
https://github.com/rainmanjam
Homepage URL
https://github.com/rainmanjam/poka-yoke
Keywords
poka-yoke, mistake-proofing, error-prevention, guardrails, code-review, defensive-design, type-safety, footgun, reliability, zero-defects
Additional notes for reviewers
poka-yoke is 11 skills, a dependency-free hazard scanner, and installable pre-commit / CI / hook templates. It applies Shigeo Shingo's mistake-proofing method to code: prefer a device that makes a wrong action impossible or self-announcing over an instruction asking someone to avoid it.
Executables, disclosed rather than left to be found. Nothing runs on install and nothing reaches the network:
scripts/detect_hazards.py — static scanner, standard library only, runs when a skill or the user invokes itscripts/cli.py — thin dispatcher for the scannerscripts/device_registry.py — regenerates a docs table; developer and CI onlyassets/devices/claude-hooks/* — templates, not installed automatically. A PreToolUse guard the user chooses to copy in. There is no hook registration in the plugin manifest
Evidence and its limits. 591 blind-graded runs across six model families, assertions written before the runs, grader blind to configuration. The behaviour it most reliably changes is stating what a design forecloses: 45% of responses did that unprompted, 80% with the skills applied, across 132 graded verdicts. That average conceals where the effect lives: asked squarely to design an interface models already do it 77% of the time, and the large gains are in tasks where nobody asked for a design review (endpoint 14% to 79%, agent feature 33% to 83%, form 29% to 64%).
Stated with its limits, because they matter: every run is the first turn of a fresh session, so this measures the ceiling rather than what survives a long working session. The baseline is no methodology rather than a different one, so it does not establish that this method in particular caused the gain — control arms exist in the repo but have not been run. And the method costs something measurable: responses became worse at spotting a raw SQL interpolation already on the page (92% to 69%) while becoming better at changing the shape that allowed it. Raw runs, the harness and the assertion checklists are in benchmarks/.
The plugin ships to 19 runtimes via native manifests; six were behaviourally benchmarked (Fable 5, Opus 5, Sonnet 5, Haiku 4.5, Codex gpt-5.6-terra, Gemini 3.1 Pro). The other thirteen are untested and docs/install.md says so in a support-tier table.
Note on keywords: plugin.json carries 24 keywords, several containing spaces for the Claude directory's own search. The list above is the hyphenated subset your validator accepts; I have not edited plugin.json to satisfy this form, since external.json carries its own keywords field.
A skill-only copy is also open as #2792. These are complementary: that entry is for discovery in the skills index, this listing tracks the full plugin at a pinned tag so updates flow from one source rather than a fork drifting. If you would rather have only one, this listing is the one to keep.
Submission checklist
Plugin name
poka-yoke
Short description
Mistake-proof your code, your pipeline, and your agents. Audits code for mistakes that are possible, designs APIs where misuse cannot be expressed, installs guardrails, and turns incidents into devices that prevent recurrence.
GitHub repository
rainmanjam/poka-yoke
Plugin path inside the repository
plugins/poka-yoke
Ref to review
v0.2.0
Commit SHA to review
18e2bc17d682ece3f13892e7415409a9da57d031
Version
0.2.0
License identifier
MIT
Author name
rainmanjam
Author URL
https://github.com/rainmanjam
Homepage URL
https://github.com/rainmanjam/poka-yoke
Keywords
poka-yoke, mistake-proofing, error-prevention, guardrails, code-review, defensive-design, type-safety, footgun, reliability, zero-defects
Additional notes for reviewers
poka-yoke is 11 skills, a dependency-free hazard scanner, and installable pre-commit / CI / hook templates. It applies Shigeo Shingo's mistake-proofing method to code: prefer a device that makes a wrong action impossible or self-announcing over an instruction asking someone to avoid it.
Executables, disclosed rather than left to be found. Nothing runs on install and nothing reaches the network:
scripts/detect_hazards.py— static scanner, standard library only, runs when a skill or the user invokes itscripts/cli.py— thin dispatcher for the scannerscripts/device_registry.py— regenerates a docs table; developer and CI onlyassets/devices/claude-hooks/*— templates, not installed automatically. APreToolUseguard the user chooses to copy in. There is no hook registration in the plugin manifestEvidence and its limits. 591 blind-graded runs across six model families, assertions written before the runs, grader blind to configuration. The behaviour it most reliably changes is stating what a design forecloses: 45% of responses did that unprompted, 80% with the skills applied, across 132 graded verdicts. That average conceals where the effect lives: asked squarely to design an interface models already do it 77% of the time, and the large gains are in tasks where nobody asked for a design review (endpoint 14% to 79%, agent feature 33% to 83%, form 29% to 64%).
Stated with its limits, because they matter: every run is the first turn of a fresh session, so this measures the ceiling rather than what survives a long working session. The baseline is no methodology rather than a different one, so it does not establish that this method in particular caused the gain — control arms exist in the repo but have not been run. And the method costs something measurable: responses became worse at spotting a raw SQL interpolation already on the page (92% to 69%) while becoming better at changing the shape that allowed it. Raw runs, the harness and the assertion checklists are in
benchmarks/.The plugin ships to 19 runtimes via native manifests; six were behaviourally benchmarked (Fable 5, Opus 5, Sonnet 5, Haiku 4.5, Codex gpt-5.6-terra, Gemini 3.1 Pro). The other thirteen are untested and
docs/install.mdsays so in a support-tier table.Note on keywords:
plugin.jsoncarries 24 keywords, several containing spaces for the Claude directory's own search. The list above is the hyphenated subset your validator accepts; I have not editedplugin.jsonto satisfy this form, sinceexternal.jsoncarries its ownkeywordsfield.A skill-only copy is also open as #2792. These are complementary: that entry is for discovery in the skills index, this listing tracks the full plugin at a pinned tag so updates flow from one source rather than a fork drifting. If you would rather have only one, this listing is the one to keep.
Submission checklist