Español

The Commands Nobody Types

Most of us settle into a handful of commands in the first week and never go looking again. That's rational — you learn what you need and get on with the work. It also means a set of built-in tools sits unopened, several of which answer questions you've been answering by hand.

/doctor — a checkup that costs nothing to run

A setup checkup that diagnoses issues and can fix them. It checks installation health — duplicate or leftover installs, PATH problems, unparseable settings files — and flags slow hooks and available updates.

The part worth the visit is the context accounting: it finds skills, MCP servers, and plugins you aren't using versus their context cost. Everything installed to try once and never removed is charged against your window on every session, and this is the only place that bill is itemised.

It also deduplicates local rules files against checked-in ones and proposes trims for checked-in rules files, cutting content the agent could derive from the codebase and keeping the pitfalls and conventions it couldn't. It reports findings first and asks before changing anything, so running it commits you to nothing.

/insights — a report on how you actually work

Generates an HTML report analysing your recent sessions on this machine: which projects you work in, how you use the tool, where things go wrong, and features to try. A single run covers up to 200 sessions it hasn't seen before, skipping very short ones.

This is a different question from /usage, which tells you what you spent. /insights is about friction — misunderstood requests, buggy output, the patterns you'd never spot from inside any single session because they're only visible in aggregate.

Reports are written to ~/.claude/usage-data/report.html, with a timestamped copy kept per run. Two caveats: the analysis runs through your normal account, so the tokens count against your plan; and it can only analyse sessions that still exist, which is governed by your retention setting. Aggressive retention quietly limits what it can see.

/branch — try a different direction without losing this one

Creates a branch of the current conversation at this point, so you can try a different direction without losing the conversation as it stands.

The usual alternative is worse in both directions: either you commit to the new approach and can't get back cleanly, or you start a fresh session and re-explain everything. Branching gives you the same context twice, going different ways — useful when you're about to try something you're only 60% sure about.

/btw and /subtask — two ways to ask without polluting

/btw asks a side question about the current session without adding to the conversation. Small, and it removes a real cost: the tangential question that permanently joins your context and travels with every subsequent request.

/subtask hands a side task to a subagent that reports back into this conversation. The distinction is scope — /btw is a question, /subtask is work. Used as a fork it inherits the full conversation and shares the parent's prompt cache, so the exploratory work stays out of your main context while still having access to all of it.

Why these stay unused

None of these are advanced. They ship in the box and most take one line to try. They go unused because nothing surfaces them at the moment they'd help: you don't think to run a setup checkup, you think "this feels slow". You don't think to branch, you think "I don't want to lose this".

A reasonable habit is to run /doctor and /insights once a month — both are read-first and report before acting — and to remember /branch exists at the specific moment you're hesitating to try something.

The takeaway

You're not out of features, you're out of commands you've tried. /doctor tells you what your setup is costing you, /insights tells you where you keep losing time, and /branch means never having to choose between trying something and keeping what you have.