The Claude Code tips pack is a shipped set of 44 situations that teach Claude Code inside Claude Code. Each one pairs the moment a person is in, in their own words, with the advice for it and the command it names. Write ok this is a mess, let's start over on the auth stuff, pause, and a note appears under the line: Starting over? /clear wipes the conversation, CLAUDE.md stays. Press underscore and the draft becomes /clear; press it again and your sentence is back.
What it teaches
The pack covers Claude Code's own vocabulary across slash commands, keybindings, flags and features: session control, context, undo, model choice, deeper reasoning, worktrees, plan mode, hooks, MCP, skills, CLAUDE.md and more. You never have to type a command's name to find it; you describe what you are trying to do. Real entries, with the note verbatim:
| You write | The note says | Underscore gives |
|---|---|---|
| ok this is a mess, let's start over on the auth stuff | Starting over? /clear wipes the conversation, CLAUDE.md stays | /clear |
| it forgot the plan we agreed on and is wandering | Losing the thread? /compact summarises the session; add a focus to say what to keep | /compact |
| that broke everything, go back to before the refactor | Undo what it just did: /rewind (or Esc twice) restores code and conversation to a checkpoint | /rewind |
| i keep having to approve every single git command | Tired of approving? /permissions allow rules like Bash(git *), or Shift+Tab to auto mode | /permissions |
| how much has this session cost so far | Wondering what this costs? /cost shows the totals; the whole history is re-sent every turn | /cost |
| i want the formatter to run automatically after every edit | Want it automatic on every edit? A PostToolUse hook runs the formatter; see /hooks | /hooks |
| make a plan before you edit anything, this touches three services | A change across several files? Ask for a plan first: Shift+Tab twice, approve, then let it edit | Advice only; underscore dismisses |
Behind each note is a situation line the model matches against, written the way a person would put it. The /rewind entry's is: wants to undo or take back what the ASSISTANT just changed or did (not an undo inside their own code, database or app). The situation decides when the note appears; the note and the command are the pack's own words, never the model's.
Why teach a tool inside the tool
This pack is the working demonstration of continuous onboarding: agentic tools ship features faster than documentation can describe them, so the teaching surface moves into the prompt itself. You learn /rewind at the moment you need it, while asking for exactly the thing it does, without knowing its name. When Claude Code's behaviour changes, updating one file updates every user's tips the next time they pause. There is no separate documentation to go stale.
How it's scoped and gated
The pack declares on-host: [claude-code], so its situations are only matched inside Claude Code; sibling packs exist for OpenCode (37 situations), Gemini CLI (41) and the shell app (15). Display is gated by tips-mode: semantic, the default, runs the match after every pause through your configured cues model; off hides every tip. Once every command is second nature, turning tips off is the intended lifecycle, and flipping them back on after a feature ships is one line. How reliably a note fires depends on that model: on a bench of 20 phrasings that should fire and 10 that should not, gpt-oss-120b flagged 20 of 20 with no false alarms and qwen-3.8-27b flagged 18 of 20 with none. Terse phrasings (that broke everything, undo) can go silent on a small model and fire on a larger one; the model is the lever, not the pack.
Make your own pack
It's one CUE.md file with a json block, so the same shape works for your team's internal CLI, your project's jargon or your own memory joggers. Each entry needs a tip, a when (the situation, in the words someone would use) and a say (the advice, naming the command). Copy the folder, write the situations, scope it with on-host:, and it hot-reloads into the host it targets. A large catalogue, such as a project pack on top of a shipped one, is split into parallel model calls of tips-shard-size (default 50). The catalogue shape is part of the open standard, version 0.12.
Related: What is tips mode? · Continuous onboarding · OpenCues in Claude Code · The claude-status blank · GitHub repo