/skill-builder · Meta

Skill Builder

The principles of a Claude skill that triggers and stays focused.

FR EN

"My /deploy skill never triggers automatically, even when I talk about deployments."

The skill audits the existing description, pinpoints the problem (trigger phrases too vague or missing, no "NOT for" clause), proposes a rewrite of the "Use when" block with real trigger phrases, and explains which of the five mechanisms (skill, CLAUDE.md, sub-agent, hook, MCP) is the right fit here.

A.

What it does

Five primitives, one right choice: skill, CLAUDE.md, subagent, hook or MCP. The heart of a skill is its description (the trigger): "Use when", real trigger phrases, and a "NOT for" clause that prevents false positives. A skimmable body (detail pushed into reference/), then packaging into a plugin + marketplace with claude plugin validate.

B.

When to use it

When you design, review or debug a skill ("why doesn't it trigger", "skill or hook?", "is my description good?"). For interactive scaffolding from scratch, use Anthropic's official Skill Creator.

The stance

The stance: a skill only helps if it triggers at the right moment AND stays silent otherwise, and the description is what decides. Anthropic provides the interactive creation tool (Skill Creator); this skill is the complementary judgment layer: the principles, the review, the packaging.

Install this skill

claude code
$ /plugin marketplace add ohugonnot/claude-skills
$ /plugin install skill-builder@web-developpeur-skills
#skills#skill-md#plugin#marketplace#meta#authoring#claude-code

Frequently asked questions

What's the difference between this skill and Anthropic's official Skill Creator?

The official Skill Creator guides interactive scaffolding from scratch (structure, files, packaging). This skill is the complementary judgment layer: auditing an existing description, debugging a failed trigger, choosing between the five available primitives.

How do I know whether to write a skill or just add an instruction to CLAUDE.md?

If the behavior should activate on a specific context and stay silent otherwise, it's a skill. If it's a permanent rule (code style, commit convention), it goes in CLAUDE.md. The skill helps you decide when it's ambiguous.

My skill fires too often (false positives). How do I fix that?

It's almost always a missing "NOT for" or an overly broad description. The skill audits your trigger phrases, identifies what's generating the false positives, and proposes a targeted "NOT for" to constrain it.

See the skill source: SKILL.md on GitHub. The skill also triggers automatically when your request matches its description, and an AI agent discovers it via the MCP server. To understand how these skills are designed, read the official skills patterns.

Feedback & discussion · 0

No feedback yet. Tried it? Tell us what you think.