Claude power words and Claude Code commands

How to use these

You do not always need a long prompt to change the kind of answer Claude gives you. Often one word is enough to name the kind of thinking you want.

Most of these are not hidden commands. They work because they name frameworks and behaviours Claude already understands, which means one word can replace a paragraph of explaining what you want. Add them to the end of a prompt, use them as a follow up after Claude answers, or combine several.

RED TEAM this.
Run a PREMORTEM.
Make this SPOKEN.
Find the BLINDSPOTS.

The first section is the exception. Those are real Claude Code commands, not shorthand. They do specific things and they only work inside Claude Code.

Real Claude Code commands

These are different from everything else on this page. They are actual features, they do specific things, and they only work inside Claude Code.

• ULTRATHINK Asks for deeper reasoning on that turn only, without changing your session effort setting. It is the only keyword of its kind. Think, think hard and think more are read as ordinary text and do nothing. Try: ULTRATHINK before deciding how to restructure this.
• /plan Enters plan mode so Claude designs the approach before touching anything. Use it when the change spans several files, or when a wrong turn is expensive. Try: /plan rebuild the authentication flow
• /batch Breaks a codebase-wide change into independent units, gives each one its own subagent in its own worktree, and opens a pull request per unit. Needs a git repository. Try: /batch migrate src/ from JavaScript to TypeScript
• /compact Summarises a long conversation to free up context while keeping what matters. You can tell it what to keep. Try: /compact focus on decisions, files changed and what is left
• /simplify Runs a cleanup only review of what you just changed, for reuse, quality and efficiency. It no longer hunts for bugs, so use /code-review when you want it finding actual problems. Try: /simplify focus on duplicated logic
• /context Shows how your context window is being used, as a grid, with warnings when you are near the limit. Run it before you hit problems, not after.
• /clear Starts a fresh conversation with empty context while keeping your project setup. Use it when you switch to something unrelated.
• /effort Sets how much reasoning Claude applies, from low up to max. Trade speed and cost against depth. Try: /effort high
• /rewind Rolls the conversation and the code back to an earlier checkpoint. This is your undo when Claude has taken the project somewhere wrong.
• /security-review Checks your pending changes specifically for security vulnerabilities. Worth running on anything touching auth, permissions, user input or APIs.

One caveat. Claude Code ships updates constantly and commands get renamed. Type /help in a session and you get the list that is actually installed in your version. Trust that over any printed list, including this one.

Three more Claude Code commands worth knowing

• /diff Opens a diff viewer showing your uncommitted changes and what Claude changed each turn. Run it before you commit, not after.
• /btw Asks a side question without adding it to the conversation history, so a quick lookup does not clutter the main thread. Try: /btw what does this env var do again
• /goal Sets a condition and keeps Claude working across turns until it is met, instead of stopping after each reply. Try: /goal all tests pass and the build is clean

Make Claude challenge you

This is the section that earns its keep. Claude defaults to being agreeable. These are how you stop it.

• RED TEAM Attacks your idea the way someone whose job is to find weaknesses would. Try: RED TEAM this strategy.
• PREMORTEM Assumes the plan has already failed and works backwards to the likely reasons. Try: Run a PREMORTEM. It is six months from now and this failed. Why?
• STEELMAN Builds the strongest possible version of the opposing argument, not a weak version it can knock over. Try: STEELMAN the argument against my position.
• BLINDSPOTS Names the important things you have not considered. Try: What are my BLINDSPOTS?
• ASSUMPTIONS Separates what you actually know from what you are treating as known. Try: List the ASSUMPTIONS this plan depends on.
• FALSIFY Looks for evidence that would show the idea is wrong. Try: Try to FALSIFY this idea.
• SKEPTIC Doubts the claims. Asks what the evidence for each one actually is.
• DEVIL’S ADVOCATE Argues against your position even when the first answer supported you.
• CRITIC Judges the execution. Reviews the work the way an experienced practitioner would, looking for weak and generic thinking. Try: Be a CRITIC before rewriting this.
• KILL CRITERIA Defines in advance what would make you stop, abandon or rethink it. Try: Set KILL CRITERIA for this before I start.

Think better

• FIRST PRINCIPLES Strips the problem back to what is actually true and rebuilds from there, instead of repeating conventional advice. Try: Use FIRST PRINCIPLES to rethink this strategy.
• COUNTERFACTUAL Changes one important assumption and shows how differently it plays out. Try: Run a COUNTERFACTUAL where our budget is half this.
• SECOND ORDER Looks past the immediate consequence and asks what happens after that. Try: What are the SECOND ORDER effects of this decision?
• INVERSION Works the problem backwards. Instead of how to succeed, it asks how you would guarantee failure. Try: Use INVERSION. How would I guarantee this launch fails?
• MECE Splits a messy problem into categories that do not overlap and do not leave gaps. Short for mutually exclusive, collectively exhaustive.
• FIVE WHYS Keeps asking why to get past the visible symptom and reach the actual cause.
• SYSTEMS THINKING Looks at how the parts affect each other rather than treating each problem on its own.
• BOTTLENECK Finds the one constraint holding the whole thing back. Try: Find the BOTTLENECK stopping this from growing.
• LEVERAGE Finds the smallest action capable of creating the biggest result. Try: What is the highest LEVERAGE move here?
• BASE RATES Grounds the answer in what usually happens in comparable situations, not the best case you have in mind.

Make better decisions

• TRUE COST Looks past the headline price to time, complexity, risk and what else you could have done. Try: Calculate the TRUE COST of each option.
• REVERSIBLE Sorts the decisions you can undo cheaply from the ones that deserve more caution. Try: Which parts of this decision are REVERSIBLE?
• OPPORTUNITY COST Shows what you give up by choosing this instead of something else.
• DECISION MATRIX Scores every option against the criteria that actually matter to you. Try: Build a DECISION MATRIX for these four options.
• 80/20 Finds the small number of actions likely to produce most of the result.
• TRADE-OFFS Makes the downside of each option explicit instead of selling you all of them. Try: Show me the real TRADE-OFFS.
• EXPECTED VALUE Compares options using both the payoff and the likelihood.
• REGRET MINIMISATION Asks which choice your future self is least likely to regret.
• SANITY CHECK A fast pass for obvious flaws, unrealistic assumptions and numbers that do not add up. Try: SANITY CHECK this before I send it.
• NEXT STEP Ends the analysis and forces one concrete action. Try: Enough analysis. NEXT STEP.

Make the writing sound human

• SPECIFICS Replaces vague claims with concrete numbers, names and examples. This is the single biggest tell that something was written by AI. Try: Find every vague claim and replace it with SPECIFICS.
• REWRITE3 Asks for three meaningfully different versions rather than three tiny wording variations. Try: REWRITE3. Make each version use a different angle.
• SPOKEN Rewrites for how it sounds said out loud rather than how it looks on a page. Best for video scripts, presentations and calls.
• TRIM Cuts words without losing the message. Try: TRIM this by 30% without losing anything important.
• PLAIN Removes jargon, corporate language and unnecessarily complicated wording.
• PUNCH Shorter, sharper sentences that are harder to skim past.
• VOICE Matches the characteristics of writing samples you provide. Give it two or three real examples first.
• NO FLUFF Kills the throat clearing intro, the filler and the generic conclusion. Try: NO FLUFF. Start with the useful part.
• HOOKS Generates opening lines only, instead of rewriting the whole thing. Try: Give me 10 completely different HOOKS.
• CONTRAST Makes a point land by showing the difference: before and after, old and new, wrong and right.

The real power move is stacking them

One word is useful. Four in sequence is a different thing, because each one operates on the output of the last.

Use FIRST PRINCIPLES to analyse this idea. Then RED TEAM it,
run a PREMORTEM, identify my biggest BLINDSPOTS, and finish
with the highest LEVERAGE NEXT STEP.

Identify the TRUE COST and OPPORTUNITY COST of each option.
Build a DECISION MATRIX. Show me the TRADE-OFFS. Then tell me
which parts of this are REVERSIBLE.

CRITIC this first. Find anything generic, clichéd or AI
written. Then make it SPOKEN, TRIM it by 25%, use PLAIN
language, and give it more PUNCH.

Run a PREMORTEM. RED TEAM the plan. List the ASSUMPTIONS and
BLINDSPOTS. Define KILL CRITERIA. Then give me the 80/20
actions to focus on first.

SANITY CHECK my reasoning. Show me the real TRADE-OFFS. Tell
me what is REVERSIBLE. Then give me one NEXT STEP to take
today.

/plan

ULTRATHINK. Before writing anything, RED TEAM your own approach
and list the ASSUMPTIONS it depends on. Then give me the plan.

Quick reference

• You want Claude to challenge you. RED TEAM
• You are worried the plan might fail. PREMORTEM
• You want the strongest counter argument. STEELMAN
• You think you are missing something. BLINDSPOTS
• You want to rethink from scratch. FIRST PRINCIPLES
• You are choosing between options. DECISION MATRIX
• You want the biggest impact for least effort. LEVERAGE, or 80/20
• Your writing sounds like AI. SPECIFICS, then SPOKEN, then TRIM
• You need better social content. HOOKS, PUNCH, SPOKEN
• A Claude Code task is complicated. ULTRATHINK and /plan
• A Claude Code session is getting huge. /context, then /compact

These are not magic spells

A power word steers Claude in a useful direction. It does not replace context. Ask it to run a PREMORTEM on a launch it knows nothing about and you get objections that could apply to anyone.

The pattern that works is context, then task, then power word.

INSTEAD OF THIS
PREMORTEM.

TRY THIS
I am launching a paid community for beginner marketers learning
AI. Here is the offer and the launch plan. Run a PREMORTEM and
give me the five most likely reasons this fails.

Short prompts can be powerful. Specific short prompts are much better.

Common questions

• Do Claude power words work in ChatGPT and Gemini? The forty shorthand words do, because they name frameworks the models already understand rather than triggering a feature. The Claude Code commands do not. Those are real features of Claude Code and work nowhere else.
• What does ULTRATHINK do? It asks for deeper reasoning on that turn only, without changing your session effort setting. It is the only keyword of its kind in Claude Code. Think, think hard and think more are read as ordinary text and do nothing.
• Are these hidden or secret Claude commands? The power words are not. They are shorthand for frameworks and behaviours Claude already understands, which is why one word can replace a paragraph. The slash commands are real, documented Claude Code features.
• How do I see which Claude Code commands I actually have? Type /help in a session. Claude Code ships updates constantly and commands get renamed, so the list installed in your version is the one to trust over any printed list.
• Why does Claude keep agreeing with me? It defaults to being agreeable. RED TEAM, PREMORTEM, STEELMAN, BLINDSPOTS and DEVIL’S ADVOCATE are the words that stop it. Use them as a follow up after the first answer so you get the contrast.
• Can I combine several power words in one prompt? Yes, and it is the biggest gain available here. Each word operates on the output of the last, so four in sequence produces something different from four separate prompts. See the stacking recipes above.