So, can you actually trust Claude to check its own work?

@itsmariahbrunner

49.2K followers · usually 2.7K views a post

Views

25.9K

Likes

1.4K

Comments

48

Shares

357

SoundTikTok sound · original sound - itsmariahbrunner

Promoted product

The AI Challengelearnaiwithmariah.com

Free AI guides, tutorials, and the Weekend Claude Bootcamp by Mariah Brunner.

8 videos

Why it worked

The video provides immediate, high-value utility by solving a common frustration with AI, which builds trust before transitioning into a relevant, low-friction pitch for the creator's course.

Summary

The creator explains why asking an AI to proofread its own work is ineffective and demonstrates a 'subagent' hack to get more accurate results. She then pivots to promoting her AI challenge course as a place to learn more advanced AI techniques.

Structure

1Identify the problem of AI self-correction
2Explain the subagent solution
3Demonstrate the multi-agent hack
4

Original caption

When you ask the same chat that wrote something to check its own work, it just agrees with itself. The fix is to spawn a subagent. Fresh chat, zero memory, actual fresh eyes on your work!! I go way deeper on hacks like this in my AI challenge. Added to my bio if you want to check it out! Or just comment “challenge”. Save this so your AI stops being an echo chamber!! #aiforbeginners #aihacks #howtouseai #claude #anthropic

More from @itsmariahbrunner

Product placement

The video uses Claude as a tool to demonstrate the hack, but the promoted product is the creator's AI course.

Call to action

It's in my bio if you're interested.

On-screen text

Claude can clone itself to catch mistakes Can you check this for mistakes? I reviewed it carefully. Everything looks good! Let me review this and check for any issues. I think we're aligned on the next steps. Let me know if you have any questions. Spawn Subagent Blank memory Spawning subagent... Here are the issues found: Incorrect statistic on line 8 Missing source for claim on line 8 Logical gap in conclusion Agent Subagent Same power. Different context. Spawn three subagents. Subagent 1: Skeptical Customer Subagent 2: Witty Copywriter Subagent 3: Sales Expert Build AI Systems That Run Your Work, Business, And Life AI agrees with you Real feedback makes you better Never let your AI become an echo chamber.

Transcript

So, can you actually trust Claude to check its own work? The short answer is no. But the fix is stupidly simple, so let me show you. So when you ask the same chat to go back and check something, it's reading its own work that has the same blind spots that gave you that mistake in the first place. It's like proofreading your own email at midnight. You're gonna miss the typo. The fix is to spawn what's called a subagent, and it opens a brand new chat with a totally blank memory, which is the whole point. It's fresh eyes, so it's like handing your work to a co-worker who's never seen it before. You literally just say, spawn this to a subagent and tell me what's wrong, and it will catch what the first one swore wasn't there. And don't let the name fool you, a subagent doesn't mean that it's less capable, it just means another agent. But here's the hack, you're not just stuck with one. I'll often tell Claude, spawn me three subagents, one acting like a skeptical customer, one like a witty copywriter, and one like a sales expert, and have each one give me their top three changes. Now I'll actually get three honest opinions and no one can see each other's answers. And this is the exact kind of hack that I go very deep on in my AI challenge. It's in my bio if you're interested. But it all comes back to one rule, do not let AI be your echo chamber.