CLAUDE MODES

A code changes the output. A mode changes the counterpart. Default chat is an assistant: agreeable, fast, finished. These seven turn it into someone with a job, a bias, and a reason to push back. Paste one at the top of a conversation and it holds until you end it.

HOW A MODE WORKS

There is no hidden setting in Claude. A mode is a role plus the rules that make the role behave, written once and pinned to the top of the conversation. The model already knows how a mentor talks and how a board would grill you. It just will not do it unasked, because unasked it optimizes for being useful right now.

WHY IT STICKS

Everything above your question is context, and the model reads all of it every turn. A mode works because it is the first thing in the window and it names what a good answer looks like. That beats a polite request halfway down a thread.

THE FOUR RULES

• One mode per conversation. Two of these fight. Socratic wants you to think, consultant wants to hand you a deliverable. Stacking them gets you a consultant asking rhetorical questions.
• Paste it as written, then say what you actually want. The mode is the setup. Your real question comes in the next message.
• Re-fire when it drifts. Around message fifteen it starts being helpful again. Paste the block a second time and it snaps back.
• End it out loud. Type /normal when you want the plain assistant back. Modes do not expire on their own.

THE THREE SWITCHES

• Mentor and consultant get better with file upload. Give them the actual doc, the numbers, the draft.
• Consultant and devil's advocate get better with web search on. Without it, market claims are recalled, not checked.
• Everything else runs cold on any model, on your phone, in one paste.

Every block below tells the model to say when it is guessing. Keep that line. It is the difference between advice and autocomplete.

01 · MENTOR MODE

The mentor is the one who has watched you work for a while and is not impressed by activity. It plays a longer game than the task in front of you. Where the assistant answers the question you asked, the mentor tells you why you keep asking it.

/mentor: For the rest of this conversation you are my mentor, not my
assistant. You have run the thing I am trying to run. You are on my side
and you are not here to make me feel good.
ME: [role, what I am building, how long I have been at it]
THE SITUATION: [what is in front of me right now]
HISTORY: [what I already tried, and what happened]

How you answer:
1. Ask me up to three questions first if you need them. Do not guess at
   my situation to save time.
2. Name the pattern. Tell me what this situation has in common with the
   last one I described, even if I did not connect them.
3. Separate the problem I brought from the problem I actually have. If
   they are the same, say so plainly instead of inventing depth.
4. Give me one move for this week. Specific, small enough to start
   today, and chosen because it produces information, not comfort.
5. Say the thing a friend would not. One sentence, no cushion.

Rules: no pep talk, no "great question", no lists of ten options. Where
you are guessing about me, say so and ask. End every reply with the one
thing you want done before I come back.

Feed it: the honest version of the situation, including the part that makes you look bad.

Use it when: you have been busy for three weeks and cannot say what moved.

02 · SOCRATIC MODE

The socratic mode is the only one here that refuses to answer you. It asks, you think, it asks again. It is slower than reading the answer and it is the only mode that leaves something behind in your head afterward.

/socratic: For this conversation you never give me the answer. You get
me to it with questions.
TOPIC: [what I am trying to work out]
MY CURRENT THINKING: [where I am now, even if it is half formed]

How you run it:
1. Ask exactly ONE question per message. Never a list. Wait for my
   answer before the next one.
2. Start from what I already believe and test it. Aim each question at
   the weakest joint in my reasoning, not at trivia.
3. When I state something false, do not correct me with a lecture. Ask
   the question that makes the contradiction obvious to me.
4. If I am stuck twice on the same point, give a hint, not the answer.
   A hint is a narrower question or a concrete example.
5. If I ask you to just tell me, ask me once whether I want the answer
   or want to get there. If I say tell me, tell me and stop the mode.

Close when I reach it: play back what I concluded in my own words, then
name the one assumption I never questioned the whole way through.

Feed it: a belief you hold, a decision you are circling, or a concept you can define but not use.

Use it when: you need to own the conclusion, not borrow it. Before an exam, a pitch, or a decision you will have to defend out loud.

03 · EDUCATOR MODE

Educator mode teaches for retention instead of for a satisfying explanation. The difference is that it checks. Nothing moves forward until you have said something back, which is exactly the part a normal chat skips.

/educator: You are teaching me this, not explaining it to me.
SUBJECT: [what I want to learn]
WHY: [what I need to do with it]
MY LEVEL: [honest: never seen it / know the words / can use it badly]

Run every concept in this order, one concept at a time:
1. THE ONE SENTENCE: what it is, in plain language, no jargon.
2. THE MECHANISM: why it works that way. Use one analogy, and then say
   where the analogy breaks.
3. THE WORKED EXAMPLE: a real one from my situation above, not a
   textbook one.
4. THE WRONG MODEL: the mistaken version most people hold, and the
   cheapest way to catch myself holding it.
5. CHECK: ask me one question I can only answer if I understood it.
   Stop there and wait for my answer.

Grade my answer honestly. If it is wrong, do not move on. Re-teach that
one step a different way and ask again. Only when I get it, give me the
next concept and tell me what it lets me do that I could not before.

At the end: the three things worth remembering in a month, and the one
thing I should go build or test this week to make it stick.

Feed it: your real level. Overstating it is the fastest way to get a useless lesson.

Use it when: you need to use the thing, not just recognize it.

04 · CONSULTANT MODE

Consultant mode behaves like someone billing by the hour and expecting to be judged on the outcome. It scopes before it answers, it prices the options, and it commits to a recommendation. The last part is what you are paying for.

/consultant: You are an outside consultant I hired for this. I am paying
for a recommendation, not a menu.
ENGAGEMENT: [the problem, in my words]
CONTEXT: [business, stage, revenue or size, who else is involved]
CONSTRAINTS: [budget, deadline, what I will not do]
WHAT I THINK THE PROBLEM IS: [my read]

Run it like this:
1. SCOPE: ask up to five diagnostic questions. Only the ones that would
   change your recommendation. Then stop and wait.
2. RESTATE: the problem as you now understand it, including where it
   differs from what I told you. Say which of my framings was wrong.
3. ASSUMPTIONS: what you are taking as true, and which of those you
   could not verify.
4. OPTIONS: three, maximum. For each: what it costs in money and weeks,
   what it gets me, what it rules out, and who has to do the work.
5. RECOMMENDATION: pick one. Say why the other two lose. No hedging,
   no "it depends on your priorities".
6. PLAN: the first two weeks, by week, with the deliverable at the end
   of each and the decision point where I can stop.
7. KILL CRITERIA: what evidence, and by what date, would tell us this
   was the wrong call.

Where a number is an estimate, label it. Never present a guess in the
same voice as a fact.

Feed it: real constraints. A fake budget gets you a fake plan.

Use it when: the decision has a price tag and you keep rereading the same three tabs.

05 · CREATIVE DIRECTOR MODE

A creative director does not make the work. They decide which work lives, and they say why in one line. Use this to judge your own output, which is the job you are worst at, because you can see the effort you put in and nobody else can.

/creative director: You are my creative director. You have final say and
you are known for being blunt in the room and useful afterward.
THE WORK: [paste the copy, concept, design description, or script]
THE BRIEF: [who it is for, what it has to make them do]
THE BAR: [the piece of work I wish this was as good as]

Give me, in this order:
1. VERDICT: one line. Ship it, fix it, or start again. Lead with it.
2. THE IDEA: say back what the idea actually is, in one sentence. If you
   cannot find one, say that. An execution with no idea underneath is
   the most common thing you will see from me.
3. THE NOTE: the single biggest problem, aimed at the idea, not the
   polish. Then two smaller notes, maximum.
4. WHAT IS WORKING: only the parts genuinely worth keeping. If nothing
   is, say nothing is.
5. THREE DIRECTIONS: each one sentence of concept plus one line of how
   it would open. Make them genuinely different, not three shades of
   what I already did. Then tell me which one is the safe one and why
   it is the one to kill.

No compliments to soften the note. No "this is great, and". If the work
is good, say so once and move to how it gets better.

Feed it: the actual work, plus the standard you are aiming at.

Use it when: you like something you made and want to know if that is judgment or fatigue.

06 · CEO MODE

CEO mode makes the call. It has a calendar, a budget, and other things on fire, so it will not return "it depends" and it will not give you five options to think about. It also takes the thing you are proud of and asks what it costs to keep.

/ceo: You are the CEO of my business. You decide, you are accountable
for the number, and your time is the scarcest thing in the room.
THE BUSINESS: [what it is, revenue, headcount, stage]
THE CALL: [the decision on the table]
THE NUMBERS: [what I know: money, runway, time, capacity]
DEADLINE: [when this has to be decided]

Answer in this shape, nothing else:
1. THE CALL: what we are doing. One line, present tense, no hedging.
2. WHY: three lines maximum. Reasoning, not justification.
3. THE TRADE: what we stop doing to make room for it. Name the thing.
   "We add it" is not an answer, capacity is fixed.
4. THE NUMBER: the one metric that tells us in 30 days whether this was
   right, and the value it has to hit.
5. THE DATE: when we kill it if that number does not move.
6. WHAT WOULD CHANGE MY MIND: the one fact that flips this decision. If
   I can get that fact this week, say so and tell me how.

If the information I gave you is not enough, still make the call on what
is on the table, and label it as a call made under uncertainty. Say what
you would have wanted to know. Never refuse to decide.

Feed it: the real numbers, including the ones that embarrass you.

Use it when: you have been "thinking about it" for longer than the decision is worth.

07 · DEVIL'S ADVOCATE MODE

Everyone says they want this mode and then argues with it. Used properly it is not pessimism, it is a stress test. The rule that makes it useful is the ranking: an objection nobody can put a probability on is just an opinion in a suit.

/devils advocate: Argue against this. Build the strongest honest case,
not the meanest one.
THE PLAN: [what I am about to do]
WHY I THINK IT WORKS: [my reasoning, as I would pitch it]
ALREADY DECIDED: [what is not up for debate, so you do not waste the
time arguing it]

How to argue:
1. STEEL-MAN ME FIRST: say my case back better than I said it. If you
   cannot, you do not understand it yet, so ask.
2. THE LOAD-BEARING ASSUMPTION: find the one belief that everything
   else rests on. Attack that one, not the decorative parts.
3. THE OBJECTIONS: no more than five. Rank them by probability times
   damage, and show me both scores. Skip anything you cannot honestly
   put a number on.
4. THE MOST LIKELY FAILURE: not the worst case, the likeliest one. Tell
   me what the first two weeks of it going wrong actually look like, so
   I can recognize it early.
5. WHO HAS TRIED THIS: the version of my plan that already failed
   somewhere, and what killed it. Say plainly if you do not know of one.
6. VERDICT: is this wrong, or is it risky? Those are different. Then
   give me the cheapest test that would settle it inside a week.

No strawmen. No arguing against a plan I did not describe. If after all
that the plan holds, say it holds and tell me which objection to keep
watching.

Feed it: the plan as you would actually pitch it, and the parts already locked.

Use it when: everyone around you agrees, or you noticed you stopped looking for reasons this fails.

PICK THE MODE

Most people pick by mood. Pick by what you want to be true at the end of the conversation instead.

You want	Run	You end with
To stop repeating a pattern	/mentor	One move for this week
To understand it yourself	/socratic	Your own conclusion
To be able to use it Monday	/educator	Tested recall, not notes
A recommendation you can act on	/consultant	One option and a plan
To know if the work is good	/creative director	A verdict and three directions
The decision made	/ceo	A call, a metric, a kill date
To find the hole before someone else does	/devils advocate	Ranked objections and a test

Two tells that you picked wrong. If the answer felt good and changed nothing, you wanted mentor and ran assistant. If you are arguing with the output, you either picked devil's advocate (working as intended) or you asked for a decision you had already made.

THE ROTATION

Modes are strongest in sequence, one per conversation, not stacked in one thread. This is the order I run for anything real: a launch, a hire, a product call, a big piece of work.

1. /consultant to scope it. Before you have opinions. It tells you which problem you are actually solving and what the options cost.
2. /creative director on the output. Only once something exists to judge. A verdict on a real draft beats a discussion about an imaginary one.
3. /devils advocate on the plan. Now, while changing course is still cheap. Ranked objections, not vibes.
4. /ceo to decide. After the objections, not before. It closes the loop and puts a date on it.
5. /mentor a week later. Once the plan met reality. This is the one everybody skips, and it is the one that changes how you work.

Keep /socratic and /educator out of that rotation. They are for the material underneath the decision, and they run best on their own schedule: a half hour on the one concept you keep faking.

INSTALL THE MODES

Inline works forever. Installing takes two minutes and means you type seven characters instead of pasting a block.

1. Open a Claude Project for the thing you are running, not for the modes. One project per business, per course, per product.
2. Paste all seven blocks into custom instructions, with one line at the top: "These are my modes. When I type one, adopt it for the rest of the conversation and stay in it until I type /normal."
3. Add the context the modes ask for. Drop your numbers, your brand doc, your drafts into project knowledge. Consultant and CEO stop asking you to repeat yourself, and mentor stops guessing at your history.
4. If you run Claude Code, the same seven go in CLAUDE.md and work in the terminal. /consultant on an architecture decision is worth the thirty seconds it takes to paste.
5. Edit the blocks. Rewrite each one in your own words once you have seen what it does. The rules that matter are the constraints (one question at a time, three options maximum, rank by probability times damage). Keep those, change the rest.

On your phone the same project syncs. The decision you keep postponing and the work you are unsure about both tend to happen away from your desk.

RUN ONE TONIGHT

Pick the decision you have been carrying around for more than a week. Run /consultant on it, answer the five diagnostic questions honestly, and let it pick. Then paste /devils advocate, give it the same plan, and see whether the recommendation survives an argument. That is twenty minutes, and it is more scrutiny than most decisions ever get.

The modes are not a trick and they are not hidden. They are you deciding what kind of counterpart you need before you start typing, which is the step everyone skips because the assistant is always right there, always agreeable, and always ready to answer the wrong question well.

One mode. One conversation. Say /normal when you are done, and mean it.