What is the difference between a Claude Code skill and a slash command?

Custom commands have been merged into skills. A file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both create /deploy and work the same way. Skills add a directory for supporting files, frontmatter to control invocation, and the ability for Claude to load them automatically when relevant.

Where do Claude Code skills live?

Personal skills go in ~/.claude/skills/ /SKILL.md and load in all your projects on that machine. Project skills go in .claude/skills/ /SKILL.md and load in that repository; commit them so your team gets them too.

How do I stop Claude from running a skill automatically?

Set disable-model-invocation: true in the skill's frontmatter. The skill then only runs when you invoke it with /name.

When does PromptConduit suggest adopting skills?

When fewer than 10% of your prompts start with a slash command and no SKILL.md skills were invoked through the Skill tool. The editor extension requires at least 8 prompts in the session; the web Trends view requires at least 5 in the period.

Adopt skills and slash-commands

Skills and slash-commands package a repeatable workflow, such as reviewing a PR, shipping a branch, or running an audit, into one invocation with a tested, consistent prompt.

When PromptConduit flags this

PromptConduit measures two things:

Slash-command adoption rate: the share of your submitted prompts that start with a slash command (a / followed by a letter, such as /review or /deploy).
SKILL.md invocations: how many times the agent called the Skill tool to load a skill, whether you asked for it or Claude chose it.

The “Capture repeatable work as skills” tip fires when the slash-command rate is below 10% and there were zero SKILL.md invocations:

Editor extension (live session): with at least 8 prompts in the session.
Web Trends view: with at least 5 prompts in the selected period.

Using either mechanism clears the tip. A single skill invocation is enough, so the tip is aimed at sessions that use neither.

Why it matters

A long prompt you retype from memory is a little different each time and is usually missing something: a test command, a review checklist item, the house style for commit messages. That inconsistency shows up as rework and as interruptions to add the forgotten constraint.

A skill fixes the best version once. It is also cheaper on context than the alternative of putting every workflow in CLAUDE.md. Anthropic’s cost guide points out that CLAUDE.md loads at the start of every session, while skills load on demand only when invoked. Moving specialized instructions into skills keeps your base context smaller. And because skill content is appended to the conversation when invoked, Anthropic’s prompt caching docs list invoking skills and commands as an action that keeps the cache.

What to do

Find the candidates. Look back over a week of prompts. The ones you have typed three or more times, nearly word for word, are your first skills. Typical examples: “review this diff against our checklist”, “write the PR description”, “run the release steps”.

Write a minimal skill. Create .claude/skills/review-pr/SKILL.md:

---
name: review-pr
description: Review the current branch's diff for bugs, missing tests, and style issues
disable-model-invocation: true
---

Review the changes on this branch against main for $ARGUMENTS:

1. Run the test suite and report failures.
2. Flag logic errors, missing error handling, and untested branches.
3. Check naming against the conventions in CLAUDE.md.
4. Summarize findings as a prioritized list. Do not edit files.

Invoke it with /review-pr (anything after the name is passed as $ARGUMENTS). Leave out disable-model-invocation if you want Claude to load the skill automatically when its description matches the task.

Reuse existing commands. Files already in .claude/commands/ keep working and also create /name commands.

Share the good ones. Project skills in .claude/skills/ are meant to be committed, so the whole team runs the same workflow. Keep personal ones in ~/.claude/skills/.

Combine with subagents. A skill can define the shape of a task that fans out across several subagents, so a complex audit runs the same way every time.

For a step-by-step walkthrough of the format, see how to write a Claude Code skill. For how skills compare with CLAUDE.md, hooks, subagents, and MCP, see Claude Code customization: 5 ways compared.

How to check it worked

The editor extension’s coaching tab lists each SKILL.md skill and subagent invoked in the live session under Skills & subagents, with use counts and success rates. Slash-command prompts are counted toward the adoption rate rather than listed. Once a skill appears in that table, or your slash-command share reaches 10%, the tip stops firing for that session. To find candidates automatically, promptconduit skills generate --local --dry-run analyzes your local transcripts for repeated workflows and previews the skills it would write to ~/.claude/skills/. See local skills for options and /docs/tools for which assistants’ transcripts it reads. Setup for the coaching tab is in /docs/editor-extension.