37 Cheat Codes to Level Up in Claude Code

Install it and responses become terse and skimmable — "written like a Boeing instruction manual, designed so the tired mechanic cannot misread an instruction." Straight to the point, speeds up the whole workflow.

Feed it a draft and it removes the "AI tells." Unusually non-verbose — built for newer models that don't need 100 examples. Scores the draft out of 50 across dimensions (rhythm, trust, authenticity, density); if below 35, it revises. You only need a few core rules.

"Why use many token when few do trick?" Cuts output tokens while preserving meaning — a 69-token response drops to 19 tokens. Also strips tokens from text and can convert skills into invocable images. The most aggressive cost-trimmer.

Enter /config → output style → set to concise. Claude leads with results, drops the "running commentary," but keeps full detail on errors and warnings. A dialed-down built-in version of the same idea.

If the context is nearly full and you clear, all context is lost — but rewind can jump you back to the previous session. "Don't try to fix broken outputs. Rewind and pull the lever to get another output."

If you deleted a file via a shell command, rewind won't restore it — "you have to use git to get that file back." Rewind only tracks Claude's own changes.

The video claims Ctrl+A "expands resume scope" to sessions across repos. This binding does not exist in Claude Code's default keybindings — Ctrl+A is listed in the official docs only as a GNU screen prefix conflict, not an action. To resume from other repos, use /resume and pick the session from the list.

The video claims Ctrl+W shows "the history of git worktrees." No such binding exists — the docs mention ctrl+w only as a keyboard-layout example, not as an action. (In most terminals Ctrl+W deletes the previous word.) There is no built-in worktree-history shortcut.

For a big old session (>1hr, >100k tokens), Claude offers to resume from a summary instead of re-reading the whole history uncached — "rereading that conversation uncached is the most expensive single request you can do."

Rename as you go ("example-rename") so sessions show names instead of opaque IDs — makes searching and resuming the exact session you want trivial.

Instead of "find the voice profile," point at the exact file with @filename. "You know it's going to the right one."

Transport a session from claude.ai (web) into your local Claude Code. "Teleport from the web version to the desktop version."

Install it and it reads your local Claude Code logs — plus Codex, Gemini, Goose, OpenCode — and reports models, inputs, outputs, total tokens, and cost across any time period. Know exactly what you're spending.

One-line session recap ("goal was researching how businesses succeed at AI search visibility") plus suggested next actions. Perfect when you can't remember where you were.

Deep in a session on a big model (Opus 5), don't switch to a smaller model — that rebuilds your whole cache. Just append "give me a quick answer" and it skips deep reasoning for that one turn, cache untouched. "1 second, September 20th 2026, a Sunday."

Add to settings.json → true. After accepting a plan, it bins all the exploration tokens (file reads, background thinking) out of the context window. The plan survives in a separate file; only the exploration is cleared. "Planning burns a lot of tokens" — this removes the leftovers.

Claude now remembers everything in co-work alongside chats, saved to individual memory files — profiles about you, topics, areas of work, people you work with. Toggle sensitive topics on/off, and import memory from other AI providers.

Every rule costs tokens (loaded every session), so compartmentalize: give a rule a path like source/api/** and it only unwraps when Claude hits that path. "The rules folder follows symlinks too" — keep one shared rules directory and symlink it into each repo.

Write notes as <!-- HTML comments --> (multi-line OK) — Claude ignores them and spends zero tokens reading them. Reserve section headers for things that actually help Claude, not yourself.

(From Paul's Programming Notes) Don't say "summarize our progress" — it returns a human-readable doc that's useless the next morning. Say "write it as a prompt for the next session." It points at exact files, mentions what you tried that failed, and starts with where to begin. Paste it into a fresh session.

"Create a UGC script and a long-form article from my recent video, make sure they pass through humanizer" — that's three skills, and you can chain five in a single prompt. Stack Simple English + Stop Slop for combined quality.

Too many similar descriptions confuse Claude into activating the wrong skill — or none. Every skill description needs three parts: (1) one-line summary of what it does, (2) triggers (when it activates), (3) anti-triggers (when it must NOT activate).

"It makes every website look exactly the same." Instead, jump into Claude Design, create a proper design system (dashboards, landing, colors, typography, shape, pills), and nothing looks like AI slop.

In the middle of a task, invoke permissions to switch to allow/ask/auto mode or add new rules — it overrides the current session without losing context or rerunning. No need to restart when you realize it needs an API call.

For visual instructions ("move this up here"), annotate an image — "remove red circle" — and Claude makes the change directly from the annotated screenshot.

Settings → Memory → import. It gives you a prompt to paste that extracts memory from another Claude install, Codex, co-work sessions, ChatGPT — paste the result into Claude's memory. "Move between tools" fast, even if not perfect.

Now you can run a scheduled task inside a project — e.g. a copywriter task that has access to the copywriting project's exact context, without setting it up separately and pointing at the context. All accessible directly.