learnaiwithrafa
ClaudeWorkflows

The Claude Code Commands You Skip Are All the Same Command

There are over 60 built-in slash commands. Most people use four. Sort the ones you are missing by what they actually do and a pattern appears — almost all of them exist to manage one resource, and it is the one quietly wrecking your long sessions.

4 min read3 sources
  • #claude-code
  • #workflows
  • #context

Almost everyone's Claude Code vocabulary is /clear, /model, /init and /compact. There are over sixty built-in commands. That's fine — nobody needs sixty.

But group the ones people skip by what problem they solve and something jumps out: the large majority manage your context window. Not features, not integrations. Context. Which is exactly the resource that silently degrades a long session while you blame the model.

Here's the map, sorted by the symptom that should make you reach for each one.

"This session has gotten dumb"

  • /context — the one to learn first. Renders current context usage as a coloured grid, so the thing you've been guessing at becomes visible. Stop estimating.
  • /compact [instructions] — summarise the conversation to free space. The optional instructions argument is the underused half: you can tell it what to preserve.
  • /clear [name] — new conversation, project memory kept. Cheaper than compacting when the last hour is genuinely irrelevant.
  • /rewind — roll back code and conversation to a checkpoint. This one surprises people: it isn't only undo for the transcript, it's an escape from a session that has gone down a bad path with files already changed.

"I don't want to pollute this session"

  • /btw [question] — pose a question on the side, and it never enters the history. Perfect for "wait, what does this flag do" mid-task.
  • /subtask — hand a side task to a subagent. The searching, log-reading and file-scanning happens in its context, and you get back the summary.
  • /branch [name] — branch the conversation to try a different direction without losing this one.
  • /fork [prompt] — copy the conversation into a new background session.

"I want it to keep working without me"

  • /goal [condition] — set a condition and Claude works until it's met. Pairs with a check it can't argue with, like a test command.
  • /background [prompt] — detach the session as a background agent.
  • /batch <instruction> — orchestrate large-scale parallel changes.
  • /tasks — list the background work this session started.
  • /loop [interval] [prompt] — run a prompt repeatedly on a schedule.

"It keeps asking permission" / "it's burning tokens"

  • /permissions — the approval rules themselves.
  • /fewer-permission-prompts — scans your actual usage and proposes an allowlist. Better than guessing at patterns.
  • /effort [low|medium|high|xhigh|max] — the effort dial. Most tasks do not need the top of it, and this is the most direct cost lever you have.
  • /usage — token usage and cost.
  • /model — switch model and save as default.

"Something is broken"

  • /doctor — diagnose setup problems.
  • /debug [description] — turn on debug logging and troubleshoot.
  • /status — current session status.

Two worth knowing about that aren't context at all

  • /code-review [level] [--fix] — reviews your working diff for bugs, with an apply mode. Distinct from /review, which is a fast read-only pass over a GitHub PR. People mix these up constantly.
  • /security-review — checks the diff for vulnerabilities specifically.

One thing that changed

Custom commands and skills have merged. A file at .claude/commands/deploy.md and a skill at .claude/skills/deploy/SKILL.md both give you /deploy and behave the same way — existing commands/ files keep working. Skills just add options: a directory for supporting files, frontmatter controlling whether you or Claude invokes them, and automatic loading when relevant.

The practical consequence is about context again: a skill's body only loads when it's used. So reference material that would be expensive sitting in CLAUDE.md costs almost nothing as a skill. If a section of your CLAUDE.md has grown from a fact into a procedure, that's the signal to move it.

Do this today

Run /context right now, in whatever session you already have open. Look at how much of the window is gone and what's holding it.

Then pick your next big-search question and route it through /subtask instead of asking it directly. Compare the grid before and after. That's the whole argument for learning these, in one measurement.

Sources