AI field guide · Cheat sheet
← All guides
Reference

The one-page cheat sheet

Print it or keep it open. It collects the commands and working habits from all three guides.

The whole idea in one line: Claude gathers context (reads your files) → acts (edits, calculates, drafts) → verifies (checks its own work), looping until done. You delegate, describe the job well, then judge whether it's right.

What the words mean

Terminalan app where you type a computer instruction instead of clicking a button
CLAUDE.mda text file that gives Claude standing instructions for your work
Skilla saved method Claude can run again
Gitversion history for a folder, so you can return to an earlier save point
MCPa standard way to connect Claude to a live tool or data source
Hookan automatic check that runs at a set point in the work

Install once

Mac / Linuxcurl -fsSL https://claude.ai/install.sh | bash
Windowsirm https://claude.ai/install.ps1 | iex
Homebrewbrew install --cask claude-code

Then cd into a folder with real work and type claude. First run opens your browser to log in. /help lists everything.

Start a task

cd <folder>point it at real files — don't paste them in
/initdraft a starter CLAUDE.md from the folder

Manage Claude's working memory

/clearwipe working memory — every time you switch tasks
/compactlong task not done? summarise & keep going
/contextsee how full the window is
/btwdrop in an aside that won't derail the task

Think before it touches anything

Shift+Tabplan mode — explores & proposes, no edits until you approve the plan

Recover when it drifts

Escstop it mid-action; context kept, just re-brief
/rewind(or Esc Esc) roll back to a checkpoint, try again
/modelswitch models — lighter routine, heavier hard thinking

Shortcuts the Anthropic team actually use

#tell it to remember something — writes it to memory for next time
!run a computer command without leaving Claude
Ctrl+Rshow the full output Claude saw — check its working
voiceturn on Mac dictation and speak the brief

Four habits that separate value from quitting

  1. Be specific. Brief it like a sharp analyst — the precise brief gets the usable answer.
  2. /clear between tasks. A polluted session is half of all "Claude got dumber" complaints.
  3. Give it a way to check itself. "Reconcile to the £35M plan and flag if off" makes the check part of the work.
  4. Plan what matters; correct early. Corrected the same thing twice? Stop, /clear, rewrite the brief.

Where it breaks (so you trust it correctly)

  • It can be confidently wrong. Never ship a number you didn't verify.
  • Garbage in, garbage out. Give it the real file, not a screenshot.
  • Long sessions rot. Quality drops as the window fills — /clear is the fix.
  • Automation amplifies mistakes. An unwatched workflow with a weak check ships bad numbers at scale.