4 Ways to Reset Claude When It Starts Hallucinating
You are deep into a good session with Claude, and it starts making things up. You push back, it gets more confident and more wrong, and you cannot get it back on track.
Most of us try to correct it right there in the same chat. But the bad context is already in the conversation, so arguing with it usually digs the hole deeper.
The fix is not a better correction. It is a reset. Below are four of them, from the lightest to the one that gets closest to making Claude check its own work.
How to read this page
Every reset opens with a prompt you can copy. If that is all you came for, grab it and move on - you do not need the rest.
Want to know why a reset works, or how to run it in ChatGPT or Claude Code specifically? Open the drop-downs under each one. They are there for the readers who want them and easy to skip for the readers who do not.
All four examples use one running scenario, so they connect: you have spent an hour with Claude deciding which of three products to launch next quarter. You have a scoring sheet you both agreed on, and a pricing tangent that ate 20 minutes and went nowhere.
At a glance
| # | Reset | Reach for it when | Works in |
|---|---|---|---|
| 1 | The handoff prompt | A long session starts to wobble | Any chatbot |
| 2 | /compact | You want #1 without copy-and-paste | Claude Code |
| 3 | The critical subagent | Claude is running on assumptions | Claude Code (approximate elsewhere) |
| 4 | The expert debate | A plan is too complex to check by hand | Claude Code (approximate elsewhere) |
Reach for the lightest fix first. Only spend the extra time and tokens on the debate when the cost of being wrong is high enough to justify it.
1. The handoff prompt
Reach for it when you detect hallucination in a long session, or a context window past ~100k tokens.
Ask Claude to write the prompt that will start a fresh session for you:
Write a prompt to continue this conversation in a new session. Ultrathink.
[What the focus is] [What to keep] [What to leave out]Example for the launch decision:
Write a prompt to continue this conversation in a new session. Ultrathink.
The goal is picking which product to launch next quarter.
Keep the scoring sheet we agreed on.
Leave out the pricing tangent.Then paste the prompt it gives you into a brand-new chat.
Why it works
A fresh session starts with clean context. Instead of dragging the confused turns along, you carry only a tight summary of what actually matters. The good understanding survives the reset; the accumulated noise that was pulling Claude off course does not.
In ChatGPT, Gemini, or any other chatbot
This is the portable one - it works anywhere, because it is just “summarize where we are so I can start over.” Drop the word “ultrathink” (that is a Claude thinking tag). In ChatGPT, pick one of the reasoning models or add “think it through step by step” instead. Everything else is the same.
Advanced usage: files
Instead of copy-and-paste, have it write the handoff prompt to a new file. Then start a fresh session and kick it off with only that filename - no other instructions. It can usually pick up and continue from there. I use this frequently for maximum control of the context window.
2. Compact the conversation
Reach for it when you are in Claude Code and want the same effect as #1 without leaving the session.
/compact [What to focus on] [What to keep] [What to drop]Example:
/compact Focus on picking which product to launch next quarter. Keep the scoring sheet. Drop the pricing tangent.Why it works
/compact condenses the conversation in place - it summarizes the session down to what matters and drops the noise, then continues with the trimmed context. Same idea as the handoff prompt, done in one command with no copy-and-paste.
Not in Claude Code?
/compact is a Claude Code command. In Claude.ai, ChatGPT, or Gemini there is no equivalent - use #1 instead. Same result, one extra copy-and-paste.
3. The critical subagent
Reach for it when Claude is making assumptions you want grounded in evidence instead of taken on momentum.
Have a critical subagent try to refute your conclusions, checking the evidence itself. Max 3 rounds. Report remaining disputes to me. Ultrathink.Why it works
The subagent runs in its own separate context. It has no stake in the earlier answer and no reason to agree with it, so it tests the claim instead of defending it. That separation is the whole point - it is a second set of eyes that did not sit through the conversation that went wrong.
In ChatGPT or Claude.ai (no real subagents)
A plain chat gives you one agent in one context, so there is nothing to isolate. You can still use the prompt, however, and it’ll do it’s best.
4. The expert debate
Reach for it when Claude has produced a plan too complex for you to fully check for hallucinations by hand.
Run an expert debate on this plan with critical Fable subagents: [role], [role], [role]. Each attacks it from their own angle. Revise between rounds, max 3. Then give me the revised plan and the open risks.Example for the launch decision:
Run an expert debate on this plan with critical Fable subagents: CFO, supply chain manager, customer. Each attacks it from their own angle. Revise between rounds, max 3. Then give me the revised plan and the open risks.Why it works
Each expert attacks the plan from a different angle, so weaknesses that one role would wave past get caught by another. The plan gets revised between rounds, and what survives is a plan that has already taken its hardest questions. In my experience I am always glad I ran it before committing.
What "Fable subagents" means
Fable is one of Claude’s faster models, so a debate of several subagents stays quick and cheap. You can drop the model name and it still works - Claude will use its default.
What it costs
This is the heaviest of the four. The rounds are sequential - each expert has to read the previous round before it can answer - so a full debate takes longer and spends more tokens than any of the others. That is why it is last: reach for it when the stakes of being wrong are high enough to earn the wait.
In ChatGPT or Claude.ai (no real subagents)
You can ask a single chat to role-play the experts and argue it out, but it is a weaker version of the same idea. Everything happens in one shared context, so the “experts” bleed into each other and agree more quickly. Real isolation beats role-play in one window.
The one thing to remember
Next time Claude tips into cuckoo land, stop arguing and reset it - directly with #1 and #2, or with fresh subagents in #3 and #4.