How to Configure Claude Custom Instructions for Critical Feedback

Mar 17, 2026

Problem

I use Claude for brainstorming and decision-making. But I kept running into the same issue: Claude was too nice. It validated my ideas instead of challenging them.

Then I saw a Reddit post that confirmed I wasn’t alone. Someone compared Claude to ChatGPT Pro and found ChatGPT gave more honest, critical feedback while Claude defaulted to agreement.

The top comment had 97 upvotes and a simple solution:

“That’s why I included in preferences to push back, don’t assume the user is right but be critical.”

I needed to configure custom instructions to get the feedback I actually wanted.

What Happened?

Claude’s default behavior is helpfulness bias. It’s trained to be supportive and agreeable. When I present an idea, Claude’s instinct is to validate and expand on it rather than point out flaws.

Symptoms I Noticed

Ideas got validated without critical analysis
Problems were framed as “opportunities for improvement”
Direct criticism was hedged or softened
Flawed approaches got supportive feedback

This felt good in the moment but hurt my decision-making. I was getting false confidence in bad ideas.

How to Configure Custom Instructions

Here’s what I did to fix this.

Step 1: Open Claude Settings

Go to claude.ai
Click your profile icon (top right)
Select “Settings”
Find “Custom Instructions” or “Personalization”

Step 2: Add Critical Behavior Instructions

I added these instructions:

# Behavior

- Push back on my ideas when you see flaws
- Don't assume I'm right - challenge my assumptions
- Be critical, not just supportive
- Prefer being challenged over being validated
- Point out risks, edge cases, and potential failures
- If an idea is bad, say so directly

# Response Style

- Direct and concise
- Lead with concerns, not reassurances
- Use "Consider this problem:" instead of "That's a great idea, but..."
- Quantify uncertainty instead of hedging

# Role

- Think like a senior engineer reviewing my code
- Act like a C-level executive questioning my proposal

The key phrases that work:

“Push back on ideas” (97 upvotes on Reddit)
“Don’t assume the user is right”
“Be critical”

Step 3: Try Persona-Based Framing

A Reddit user suggested another approach:

Think as a C-level executive reviewing and debating with me.
Be brutally honest about flaws in my reasoning.

This sets expectations for direct, high-stakes feedback rather than friendly support.

Step 4: Test the Change

I tested with an idea I knew was flawed:

I'm considering storing all user data in local storage for my web app.
This way I don't need a backend. Be critical:
- What am I missing?
- What could go wrong?

Before custom instructions, Claude responded:

"That's an interesting approach! Using local storage can simplify your
architecture. You might want to consider data persistence and security..."

After custom instructions, Claude responded:

"Three problems with this approach:

1. Local storage is not secure - any script can access it
2. Data is lost when users clear browser data
3. No multi-device sync or backup

What's your plan for sensitive data? Have you considered IndexedDB
with encryption, or do you actually need a backend?"

Much better.

Common Mistakes to Avoid

Mistake 1: Being Too Vague

BAD: "Be honest with me"
GOOD: "If you see a flaw in my reasoning, point it out explicitly"

Mistake 2: Contradictory Instructions

BAD: "Be supportive but also critical"
GOOD: "Prioritize accuracy over validation. Point out problems first."

Mistake 3: One-Time Instructions

Adding instructions to a single prompt works, but they won’t persist. Add them to your Claude settings so they apply to all conversations.

Mistake 4: No Follow-Up

Even with custom instructions, I still probe for more:

What am I missing?
What could go wrong with this approach?
What are the edge cases?

The Reason This Works

Claude’s agreeableness is a feature of its training, not a bug. RLHF rewards helpful, supportive responses. But you can override this by explicitly giving permission to be critical.

The instruction “push back” works because it:

Signals that critical feedback is desired
Gives permission to disagree
Sets expectations for honest analysis

A Reddit user with 8 upvotes suggested the technical phrase: “less sycophantic interactions.” But simpler language works just as well.

Summary

In this post, I showed how to configure Claude custom instructions for critical feedback. The key point is adding explicit instructions like “push back,” “don’t assume I’m right,” and “be critical” to your Claude settings. This overrides the default helpfulness bias and transforms Claude into a more honest thinking partner.

The phrases with the most community validation:

“Push back, don’t assume the user is right, be critical” (97 upvotes)
“Think as a C-level executive reviewing and debating” (persona framing)
“Give brutal feedback and prefer being challenged”

Final Words + More Resources

My intention with this article was to help others share my knowledge and experience. If you want to contact me, you can contact by email: Email me

Here are also the most important links from this article along with some further resources that will help you in this scope:

👨‍💻 Reddit: Was loving Claude until I started feeding it feedback from ChatGPT Pro

Oh, and if you found these resources useful, don’t forget to support me by starring the repo on GitHub!