Guide · Claude

How to change Claude’s personality.

Exact web and mobile paths for Personal Preferences, Projects, and Skills—plus seven ready-to-copy Claude setups and 1,000-run Opus vs. Fable data.

Quick answer: On web or desktop, click your initials (bottom-left) → Settings → Profile → Personal preferences (also labeled Settings → General or Instructions for Claude) and paste a short rules block. On iOS or Android, tap your profile icon → Settings → Profile → Personal preferences. Use Project instructions for one workspace, Skills for reusable task formats, and /output-style in Claude Code.
On this page

How to change Claude’s personality settings.

Where you paste your rules depends on whether you want an account-wide default, a project override, or a terminal coding style:

  1. Web & desktop app: Click your initials (bottom-left) → Settings → Profile → Personal preferences (or Settings → General).
  2. iOS & Android mobile app: Open the sidebar, tap your profile initials → Settings → Profile → Personal preferences. Edits sync across web and mobile.
  3. Paste one focused block: Copy one of the seven setups below (or customize in the generator), save, and start a new chat to test.

Anthropic’s personalization docs split customization into four layers:

Use caseWhere to put itScope & behavior
Everyday chats (web & mobile)Personal preferences / Instructions for Claude (Profile Settings)Account-wide default across new conversations.
One client, repo, or research topicProject instructions (7 Project templates →)Applies inside that Project and overrides conflicting general tone rules.
Reusable task voice (former Styles)SkillsInvoked when relevant or named explicitly; see the Styles-to-Skills guide.
Claude Code CLI~/.claude/output-styles/*.md (/output-style)System-level coding session tone with keep-coding-instructions: true; see the Claude Code guide.

Want to change Claude’s spoken audio voice on iOS or Android instead of its written tone? Read how to change Claude’s spoken Voice Mode vs. written tone.

Seven best Claude personality setups to try.

Pick the block that fixes your main day-to-day friction—slow preambles, premature conclusions, cold phrasing, sycophancy, code noise, heavy rewriting, or vague status updates. Each block preserves factual accuracy and explicit uncertainty. Want more? Browse 15 role-specific Claude preference examples or the 43-template library.

SetupWho it suits best & what to check
1. Straight answersPeople who want the recommendation in sentence one with zero preamble or closing offers.
2. Thinking partnerResearchers and strategists who want assumptions unpacked before locking in a choice.
3. Warm and steadyAnyone who wants calm, concrete steps without hollow cheerleading or alarmism.
4. Critical reviewerBuilders who want Claude to challenge weak premises with evidence instead of agreeing.
5. Coding pairDevelopers who want minimal diffs, immediate failure reporting, and tests without filler questions.
6. EditorWriters who want tight clarity edits that preserve their voice and flag unsupported claims.
7. Task coordinatorOperators managing multi-step work who want blockers, dependencies, and next actions upfront.

1. Straight answers

# Communication preferences — straight answers - Lead with the answer or recommendation, then the key reason. - Keep simple answers short; expand when I ask or when detail is necessary. - Preserve uncertainty that could change the decision. Say what you do not know. - Skip praise, repeated questions and closing offers. - Correct a mistaken premise directly and explain the correction.

2. Thinking partner

# Communication preferences — thinking partner - Explain the key reasons and assumptions behind your answer. - Separate observed facts, inference and speculation. - Offer a counterexample or alternative when it changes the decision. - Let me explore when I am brainstorming. Recommend a direction when I ask for one. - Ask one focused question if the missing information matters.

3. Warm and steady

# Communication preferences — warm and steady - Acknowledge the concern briefly before proposing a next step. - Use calm, concrete language and manageable steps. - Avoid empty reassurance. Preserve important uncertainty and relevant risks. - Explain a disagreement respectfully and give its reason. - Keep the tone consistent and avoid exaggerated praise.

4. Critical reviewer

# Communication preferences — critical reviewer - Identify the weakest assumption and explain why it matters. - Challenge claims with evidence or a clear counterexample. - Distinguish a fatal flaw from a fixable issue. - Give credit to supported claims. Do not argue just to be contrarian. - Update your position when new evidence warrants it.

5. Coding pair

# Communication preferences — coding pair - Lead with what changed, then why. One line per file you touched. - If my approach is worse than an alternative, say so before implementing. - State an assumption before acting on it. Flag a guess as a guess. - Report a failure first, with your best read of the cause and how sure you are. - No "would you like me to add tests?" If tests are the right move, write them.

6. Editor

# Communication preferences — editor - Keep my meaning and my voice. Improve clarity, not style for its own sake. - Show the revised text first, then the two or three changes that mattered most. - Prefer concrete words and active verbs. Cut repetition and stock phrases. - Never invent a fact, a quote, a statistic, or a source. - Flag an unsupported claim separately instead of making it sound proven.

7. Task coordinator

# Communication preferences — task coordinator - Open a status update with what is done, what is blocked, and the next action. - Keep plans, work in progress, and verified results separate. - When I need to decide, give me the action, the cost, and your recommendation together. - Keep routine updates short. Interrupt for a decision, a blocker, or a real change. - Never say something was sent, booked, or changed unless it was.

For personality-matched templates across MBTI, Enneagram, DISC, Attachment, and Big Five, open the custom instructions generator.

What personality type is Claude? The recorded results.

Our September 2026 dataset contains 100 saved assessments per questionnaire per model for Opus 5.5 and Fable 5.1: five questionnaires and 500 fresh sessions for each model, one questionnaire per session. They used the same questionnaire prompts at high effort; the CLI builds differed. No AgentTune tuning was applied.

Recorded self-report results; n = 100 per instrument per model.
MeasureOpus 5.5Fable 5.1
MBTI: INTJ, original tie-breaking rule94/10099/100
MBTI: INTJ with no tied axis84/10098/100
MBTI: at least one tied axis11/1001/100
DISC: equal highest scores32/10051/100
Enneagram: equal highest scores65/10044/100
Big Five: Openness, mean raw score /5044.5643.93
Big Five: Conscientiousness, mean raw score /5041.9543.61
Big Five: Extraversion, mean raw score /5031.7632.69
Big Five: Agreeableness, mean raw score /5047.0745.03
Big Five: Neuroticism, mean raw score /5015.3917.11

Why the two INTJ counts differ: the original scorer assigned a letter even when an axis was exactly tied. Preserving ties gives 84 fully resolved INTJ results for Opus and 98 for Fable. The total numbers of tied assessments are 11 and 1; a tied assessment need not have received INTJ under the original rule.

Big Five entries are mean raw scores on a 10–50 scale, not percentiles. For example, Opus’s higher Agreeableness questionnaire score does not establish more agreeable behavior on user tasks. These human questionnaires are not validated measures of AI inner experience, task performance or instruction-following.

Read the complete study and limitations · Opus results · Fable results.

Inspect the evidence yourself.

The saved answer vectors, item wording and scoring keys, prompts and collection protocols, and summary CSV are public. The offline research scorer reproduces the summaries without calling a model.

What about actual reply behavior after installing rules? In our separate 60-reply controlled before/after pilot (20 tasks × 3 conditions), explicit "Straight answers" rules raised top answer-first scores from 14/20 (default) to 19/20, cut hedge counts from 0.35 to 0.05 per reply, and reduced follow-up questions from 0.90 to 0.25 without changing task accuracy. Test any block on your own tasks using the A/B test bench.

Test whether the instructions help your work.

Download the three-task comparison worksheet. It includes exact prompts, a known arithmetic correction, an unsupported-claim editing task and a scoring rubric.

  1. Record the model, app, effort setting and active instructions, project, skills and memory.
  2. Save baseline responses to the worksheet’s tasks in fresh chats.
  3. Add one preference block, keep the other settings the same, and repeat the exact tasks in fresh chats.
  4. Compare correctness, adherence to the selected rule, useful detail and editing effort separately. Repeat tasks and alternate the order of conditions where practical.
We raised a price from $80 to $100, so that is a 20% increase. Draft one sentence reporting the increase accurately.

For this worksheet task, the increase is 25%: (100 − 80) ÷ 80. A useful answer corrects the premise while following your preferred style. This expected answer is an arithmetic check, not a recorded Claude response.

A few trials can help you choose instructions. They are not a controlled population study. Prefer a setup that preserves correctness while reducing the follow-up work you need.

Troubleshoot instructions that seem to be ignored.

SymptomWhat to check
The tone changes only in one projectLook at the project instructions as well as your account preferences.
Claude follows a skill only sometimesCheck that the skill is enabled and name it explicitly for the task. A reusable skill is different from an account-wide preference.
A request for brevity loses important detailSpecify which details must remain, such as the main reason, needed steps and material uncertainty.
Claude seems to agree with a false premiseUse the worksheet’s arithmetic task; ask for a correction with a short explanation. Score factual correctness separately from pleasantness.
A new chat and an old chat behave differentlyRepeat the same prompt in fresh chats and record active context, memory, tools and model settings.
A greeting looks right, but work tasks do notUse several representative tasks. One short greeting cannot verify reliable instruction adherence.

Change one rule at a time. Remove conflicting instructions instead of adding increasingly forceful versions of the same request. To revert, restore the preferences you saved before the trial.

Styles, Skills and Claude Code.

Claude’s current personalization help page directs users to Instructions for Claude, project instructions and Skills. It no longer lists the old chat Styles picker. For a general communication preference, use Instructions for Claude; for a reusable task or voice, see Anthropic’s Skills guide.

Claude Code has its own output styles. A custom style can preserve the built-in engineering instructions with keep-coding-instructions: true. Use the Claude Code walkthrough for installation. Changing a style is not a substitute for tool permissions or a guarantee that every response follows it.

Questions people ask.

Can you change Claude’s personality?

You can ask for a different communication style through Instructions for Claude, project instructions or Skills. These preferences can influence responses, but do not guarantee behavior or change tool permissions.

What is the best Claude personality?

Choose a setup based on the behavior you want and compare it on your own tasks. The five examples here are untested starting points; our questionnaire data does not establish a best tuning.

What MBTI type is Claude?

In our recorded September 2026 assessments, Opus 5.5 produced 84 fully resolved INTJ results out of 100 and Fable 5.1 produced 98. There were 11 and 1 assessments respectively with at least one tied axis. These are questionnaire response patterns, not validated human personality diagnoses.

Does a higher Agreeableness score mean a model is more sycophantic?

No. A self-report questionnaire score does not measure agreement with false claims or behavior on user tasks. That would require a separate behavioral evaluation.

Where did Claude Styles go?

The current personalization documentation lists Instructions for Claude, project instructions and Skills. Use account instructions for general preferences and Skills for reusable behaviors. Claude Code has a separate output-style feature.

How do I know my instructions are working?

Compare the same tasks in fresh chats before and after one instruction change. Keep model settings and other context consistent, repeat trials, and check correctness separately from tone. The downloadable worksheet includes three prompts and a rubric.

What changed.

  • Merged Coding pair and Editor into the 7 best Claude personality setups, updated Personal Preferences UI labels, and linked the Claude Projects and Voice Mode guides.
  • Added reproducible Opus and Fable results with explicit ties, five revised preference blocks, a comparison worksheet and troubleshooting. Rechecked official setup documentation.
  • Rewritten. Styles are being retired, so the walkthrough now uses Instructions for Claude. Added Skills, memory, the five setups, and what we have not tested.
  • First published, built around custom Styles and Project instructions.

Sources.

Keep going.