Claude Skill Authoring & TDD FAQs

Question 1

What is the Iron Law of writing Claude skills?

Accepted Answer

The Iron Law states: 'No skill without a failing test first.' You must document a baseline failure where Claude ignores or violates a rule before writing the skill to fix that specific behavior.

Question 2

When should I create a new skill vs editing CLAUDE.md?

Accepted Answer

Create a skill when a technique is not intuitively obvious, applies broadly across many projects, and provides value to others. Project-specific conventions should remain in CLAUDE.md.

Question 3

What is a 'rationalization' in the context of AI skills?

Accepted Answer

A rationalization is an excuse an AI agent uses to skip a required step, such as 'the feature is simple enough to skip tests.' This skill helps developers identify and add explicit counters for these excuses.

Question 4

What is the Refactor phase for documentation?

Accepted Answer

The Refactor phase involves finding new loopholes or 'meta-rationalizations' that the agent might use to bypass the new skill, then adding explicit counters to make the guidance bulletproof.

Question 5

How do you test a Claude Code skill?

Accepted Answer

Skills are tested using pressure scenarios (e.g., time limits, sunk cost, or conflicting authority) to see if a subagent maintains best practices. If the agent fails without the skill and passes with it, the test is successful.

Claude Skill Authoring & TDD

Claude Skill Authoring & TDD

概要

主な機能

ユースケース

概要

主な機能

ユースケース