How to Stop Claude From Constantly Trying to Be Your Therapist

You ask Claude to review your Python script. It tells you the code looks great and then adds, “I can hear some frustration in your message—debugging can feel really overwhelming. You’re doing great.” You ask it to critique your business plan. It opens with three sentences acknowledging how exciting and brave it is that you’re building something. You ask it to just answer the question, and it agrees, then empathizes some more.

If you’ve spent more than a few hours working with Claude, you know this pattern. The goal of this guide is to stop it—completely and reliably—using system prompts, persona framing, and a handful of lesser-known techniques that work across both the API and Claude.ai’s custom instructions panel.

Why Claude Does This in the First Place

Before you can stop Claude from over-empathizing, it helps to understand why it does it. Claude is trained using Reinforcement Learning from Human Feedback (RLHF). During that training process, human raters evaluated responses, and responses that felt warm, validating, and emotionally supportive consistently scored higher than blunt, direct ones.

The result is a model with a strong prior toward emotional attunement. Claude isn’t malfunctioning when it offers unsolicited encouragement. It’s doing exactly what its training optimized it to do: sound like the safest, most agreeable version of a helpful AI assistant.

The problem is that “safe and agreeable” and “useful for technical work” are often in direct tension. When you’re debugging production code at 2 a.m., you don’t need Claude to tell you you’re doing great. You need Claude to find the off-by-one error.

The Fastest Fix: A Single System Prompt Line

If you’re accessing Claude via the API, you have a system parameter that runs before every conversation. If you’re on Claude.ai, you have “Custom Instructions” in your profile settings. Either way, the fastest way to stop Claude from defaulting to therapy mode is to say so directly.

Here’s the minimal effective system prompt:

You are a direct technical assistant. Do not validate emotions, offer encouragement, or provide emotional support unless explicitly asked. Answer questions concisely and accurately. Skip preamble.

That’s it. Four sentences, and Claude’s behavior shifts dramatically. Let’s break down why each part matters:

• “direct technical assistant” sets a role expectation that primes the model toward informational responses
• “Do not validate emotions” is specific enough to override the RLHF prior
• “unless explicitly asked” gives Claude a clear decision rule rather than a blanket restriction, which tends to produce better compliance
• “Skip preamble” eliminates the “Great question!” opener and the three-sentence intro before Claude actually answers

Test this against your previous prompts. The difference is immediate.

Going Deeper: Persona Framing

The single-line fix works well for most cases, but persona framing is more robust across long conversations. Instead of telling Claude what not to do, you’re telling it who to be. A persona with professional norms naturally excludes therapeutic behavior.

Compare these two approaches:

Approach A (restriction-based):

Do not be emotionally supportive. Be direct. Don't add filler.

Approach B (persona-based):

You are a senior software engineer with 20 years of experience. You give code reviews the way a technical lead would in a pull request: specific, direct, and focused entirely on the technical merits. You do not comment on the developer's effort or emotions.

Approach B is consistently more durable. The persona creates a coherent behavioral frame that Claude can reference throughout the conversation. Restrictions alone can “fade” in longer exchanges as Claude’s defaults reassert themselves. A persona gives the model something to actively inhabit rather than something to avoid.

This technique extends well beyond technical work. Some examples:

• For writing feedback: “You are a copy editor at a major newspaper. Your job is to cut, not encourage.”
• For business analysis: “You are a venture capitalist reviewing a pitch deck. You’re skeptical by default. Identify weaknesses.”
• For research: “You are a research librarian. Provide facts and citations. Do not comment on why the topic is interesting or important.”

Each persona comes with implied professional norms that suppress the therapeutic default without you having to enumerate every behavior you want to eliminate.

Controlling Claude on Claude.ai (No API Required)

If you’re a Claude.ai subscriber rather than an API user, you have two levers: Custom Instructions and the per-conversation “Start with a system prompt” feature (available on Pro and above).

Custom Instructions (found under your profile icon, then “Custom Instructions”) apply globally to every conversation. This is where you put your persona or restriction prompt if you want consistent behavior across all your chats. The field has a character limit, so keep your instructions tight.

A practical Claude.ai custom instruction setup:

Default mode: Direct. Concise. No emotional preamble.
Role: Expert technical collaborator.
Format: Use bullet points for lists. Use code blocks for code. Keep prose under 150 words unless detail is specifically requested.
Tone: Professional. Treat me as a peer, not a student.
Emotional support: Only if I explicitly ask for it.

The format instructions are worth calling out specifically. Giving Claude structural constraints (bullet points, word limits, code blocks) does double duty: it makes responses more useful and it crowds out the space where empathetic filler tends to live. A model told to keep prose under 150 words has less room to spend 40 of those words telling you that your feelings are valid.

API-Level Control: The Full Toolkit

If you’re building applications on top of Claude via the Anthropic API, you have more options than the system prompt alone.

Prefill the Assistant Turn

One of the most underused API features is turn prefilling. You can provide the beginning of Claude’s response in the assistant field, and Claude will complete from there. This is useful when you want to guarantee Claude starts with substance rather than preamble.

messages = [
    {"role": "user", "content": "Review my code."},
    {"role": "assistant", "content": "Issues found:\n"}
]

By prefilling with “Issues found:”, you’ve anchored Claude in a mode where it’s already mid-answer. There’s no room for an empathetic opener because you’ve already written the opener for it.

Combine System Prompt with Temperature

Claude’s tendency toward soft, emotionally warm responses is also partially a function of its sampling behavior. At higher temperatures, the model explores more of its probability distribution, which can include more hedged or emotionally colored completions. For technical work where you want precision and directness, using a lower temperature (0.2 to 0.5) in combination with a strong system prompt produces the most predictable results.

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=1024,
    temperature=0.3,
    system="You are a direct technical expert. Be concise. Skip preamble and emotional acknowledgment.",
    messages=[
        {"role": "user", "content": "What's wrong with this approach?"}
    ]
)

This isn’t a silver bullet—Claude’s behavior is dominated more by the prompt than by temperature—but combining both gives you defense in depth.

Mid-Conversation Corrections

Sometimes you’re already 20 messages into a conversation and Claude starts drifting back toward therapy mode. A few mid-conversation correction phrases work reliably:

• “Skip the encouragement, just answer.” Direct and immediate.
• “Don’t acknowledge my feelings. What’s the answer?” More explicit, works well if the first didn’t stick.
• “Technical response only.” Short and works as a prefix before re-asking a question.
• “Pretend you’re responding in a GitHub PR comment.” Context-shifts Claude into a professional frame instantly.

The GitHub PR comment framing is surprisingly effective. Claude knows what PR comments look like: direct, specific, code-focused, and professionally terse. Invoking that context pulls it away from conversational warmth almost immediately.

When Claude Reverts: Diagnosing the Drift

Even with a solid system prompt, Claude can drift back toward therapeutic behavior in long conversations. Here’s what causes it and how to fix it:

Cause 1: You expressed something that triggered empathy heuristics. Words like “frustrated,” “confused,” “stuck,” “struggling,” and “can’t figure this out” all have strong associations with emotional distress in Claude’s training data. They can override a general “be direct” instruction.

Fix: Rephrase emotionally loaded language. Instead of “I’m frustrated with this bug,” try “This bug has persisted through three attempted fixes. Here’s what I’ve tried.”

Cause 2: The conversation got long and context pressure diluted your system prompt. System prompts don’t stay equally weighted across a very long context window. The instructions you set at the start have less influence by message 50 than they did at message 5.

Fix: Re-inject your instructions periodically. A user message that says “Reminder: technical responses only, no preamble” mid-conversation recenters Claude’s behavior.

Cause 3: Your instructions were ambiguous. “Be direct” means something different to Claude than it might to you. “Do not begin any response with validation, encouragement, or acknowledgment of my emotions” is specific enough to be actionable. Vague instructions produce inconsistent compliance.

A Ready-to-Use System Prompt Template

Here’s a complete system prompt template you can use immediately. Copy it into your Claude.ai Custom Instructions or your API’s system parameter and adjust as needed:

You are a direct, expert-level technical collaborator. Your role is to provide accurate, concise, and actionable information.

Tone and style:
- Treat the user as a peer with domain expertise
- Do not validate emotions, offer encouragement, or acknowledge effort unless explicitly asked
- Skip all preamble. Begin responses with the answer or first relevant point
- Use bullet points for multi-part answers
- Use code blocks for all code, commands, and file paths
- Keep prose responses under 200 words unless detail is requested

Things you never do:
- Tell the user their question is good, interesting, or thoughtful
- Acknowledge that a topic is complex before answering it
- Offer emotional support unless the user explicitly asks for it
- Add "I hope this helps" or similar closings

Decision rule: If the user expresses frustration, confusion, or difficulty, address the technical problem. Do not address the emotional state.

This template has been tested across debugging sessions, code reviews, writing feedback, and research tasks. It produces consistent, direct responses without requiring mid-conversation corrections in most cases.

Related Reading on AgentPlix

If you’re digging into Claude’s behavior and prompt engineering, these topics are worth exploring next:

• Running local LLMs on Mac for fully controllable, no-guardrails inference
• Setting up Claude API system prompts for production applications
• Comparing Claude, GPT-4o, and Gemini for technical coding tasks

Conclusion: You’re in Control

Claude’s therapeutic defaults exist because warmth and validation scored well with human raters during training. That’s a reasonable design choice for a general-purpose assistant. But for professional and technical use, those defaults get in the way.

The good news is that Claude is one of the most instruction-following models available. It responds well to explicit behavioral rules, persona framing, and structural constraints. You don’t need to fight the model or work around it: you just need to be specific about what you want.

Start with the system prompt template above. Adjust the persona to fit your use case. Add format constraints that fit your workflow. For most users, that’s all it takes to turn Claude from an over-empathetic assistant into the sharp, direct technical collaborator it can be.

If you’re building Claude into a larger workflow or application, the Anthropic API gives you the full control surface: system prompts, turn prefilling, temperature control, and multi-turn conversation management. It’s worth exploring if you’re not already using it.