Engineering Verification Skill — AI skill for Claude Code

engineering-verification

A portable behavioral protocol that makes an AI coding agent work like a disciplined senior engineer inside a repository it did not write. It contains no project-specific facts; the agent discovers the stack, commands and architecture of whatever repository it is given.

The problem it solves

Coding agents tend to assume a stack, invent commands, call every red test a regression, "fix" things nobody asked about, restructure architecture that works, create a second copy of an abstraction that already exists, spread a small feature across unrelated modules, and overwrite work in progress. This protocol replaces those habits with: understand first, measure instead of guess, baseline before changing, classify failures with evidence, reuse before creating, keep changes local, and stop when uncertain.

The three modes

Mode	Use it to	Default behavior	Output
1. Investigation	understand an unfamiliar or existing repository	read, analyze, report; no code changes unless asked	templates/investigation-report.md
2. Verification	decide whether the repository is a stable baseline, or check a recent change	baseline, run the checks that exist, classify failures, fix only justified issues, re-verify	templates/verification-report.md
3. Feature safety	add a feature without collateral damage	investigate, identify the owner, reuse existing abstractions, write an impact plan, implement locally	templates/feature-impact-plan.md

It is not for redesigns, framework migrations or broad cleanups; it says so and asks what outcome you want.

Contents

engineering-verification/
├── SKILL.md              the authoritative protocol; usable without the other files
├── README.md             this file
├── references/           deeper methodology
│   ├── investigation.md      how to investigate a repository; facts vs inferences vs open questions
│   ├── verification.md       baseline rec