CockpitCockpit
← Back to blog

Slash modes in Claude Code: /qa, /fx, /review, /commit

Published April 29, 2026 · 5 min read

A subtle but huge productivity unlock: instead of writing the same "do X but don't change code" prompt every time, save it once as a slash command. Cockpit ships with four built-ins — here is what each does and when to reach for it.

One pattern keeps re-emerging in every productive Claude Code transcript I've read: the user starts by telling the agent what not to do.

"Don't change any code yet. Just walk me through the bug."

"Don't refactor. Just review the diff."

"Don't write the feature yet. Restate it back to me first."

These are postures, not tasks. They flip the agent from "ship code by default" to "be a particular kind of collaborator". And once you have the posture right, the actual conversation gets shorter and faster.

Slash commands let you save these postures as shortcuts. Cockpit ships four built-ins — same shape as Anthropic's ~/.claude/commands/ mechanism, just preinstalled.

/qa — clarify, never code

The classic "talk first, code never" mode.

/qa I want to add a soft-delete to user accounts.

The agent will:

  • Restate the requirement back in its own words.
  • Ask back on every ambiguity (cascade? grace period? reversible from UI?).
  • Follow KISS — never propose three architectures when you asked one question.
  • Refuse to write code until you say so.

Best for: spec'ing a new feature, debating an approach, onboarding into an unfamiliar repo.

/fx — diagnose, never edit

Bug evidence-chain mode.

/fx The /api/users endpoint sometimes returns 500 with no log line.

The agent will:

  • Read the failing path end-to-end.
  • Build an evidence chain: which file, which assumption, which observed behavior.
  • Refuse to edit a single file.
  • End with a concrete hypothesis you can confirm or reject.

This is the difference between "AI fixed something and broke two more things" and "AI showed me what was actually wrong, so I fixed it correctly". Use /fx when the bug surface is unfamiliar or when the cost of a wrong fix is high.

/review — read the diff, write notes

Pre-PR review without rewriting.

/review

(no argument needed — it picks up the current diff)

The agent will:

  • Read git diff and the surrounding files for context.
  • Write line-by-line review notes in the form a senior engineer would: tradeoffs, missing tests, edge cases, "why not X?".
  • Not touch your code.

The trick: feed those notes into a fresh /qa round and let yourself decide which to address. You get a second pair of eyes that doesn't have ego.

/commit — commit, in your repo's style

The least exciting one, used most.

/commit

The agent will:

  • Read your last 20 commit messages to learn your style (Conventional? Imperative? Emoji?).
  • Stage the right files (skips lockfiles, generated dirs).
  • Draft a message in your style.
  • Commit.

It's a 10-second time saver per commit, but you commit 30 times a day.

Make your own

Put any markdown file in ~/.claude/commands/ (or per-repo ./.claude/commands/) and it becomes a slash command. The file body is the system prompt. Examples that work surprisingly well:

  • /security — "Review only for security issues, no style or perf comments. List CVE-likes first."
  • /migrate — "I'm migrating this file from X to Y. Don't change behavior. Don't reformat. Show diff only."
  • /why — "Explain why this code exists. Look at git blame, surrounding tests, and any TODO/FIXME within 10 lines."

Cockpit auto-loads any file you drop in there into the slash autocomplete menu.

The bigger principle

Default Claude Code is biased toward action — write code, save files, ship. That's a great default 60% of the time and a terrible default the other 40%. Slash modes are how you flip the default deliberately.

If you find yourself typing the same "don't do X" preamble more than twice in a week, that's a slash command waiting to happen.

npm i -g @surething/cockpit · GitHub · Try Online