How to Branch an AI Conversation Without Losing the Original
When and how to branch (fork) an AI chat from any response: practical branching patterns, prompts to use after the fork, and how Fork from here works on magicdoor.ai.
Direct answer: Branching an AI conversation means copying a chat up to a chosen answer and continuing in a new thread, so you can try a different direction while the original stays intact. In magicdoor.ai, select Fork from here under any assistant response in a saved chat. You get a new chat with every message up to that point, the same skill and project, and copies of the attached files.
Fact-checked November 16, 2026 against the current magicdoor.ai chat code. This page owns branching a conversation. To score two models on the same message, see How to compare AI models side by side; for long-running project context, see Project instructions best practices.
Why branch instead of scrolling back
Long chats drift. You ask a follow-up, the model takes it literally, and ten messages later the thread is optimized for a direction you no longer want. Editing your way back is messy, and starting a fresh chat throws away the context you spent time building.
A branch keeps the useful setup and gives you a clean fork in the road:
- Explore two directions from the same brief without the second one being influenced by the first.
- Protect a good answer before you push the model harder.
- Hand off a clean thread without the dead ends.
- Switch models from a known state, so the new model starts with the same context the old one had.
Five branching patterns that actually help
| Pattern | Fork at | Then send | Why it works |
|---|---|---|---|
| A/B direction | The answer where you agreed on the brief | "Take direction B: [describe]. Keep everything else from the brief." | Each direction builds on the same agreed brief instead of on the other direction |
| Save point | The best draft so far | "Make it 30% shorter and more direct. Keep every factual claim." | If the edit goes wrong, the original thread still holds the good draft |
| Model handoff | The answer that set up the problem | Switch the model, then: "Continue from here. Check the plan above for mistakes before extending it." | The second model reviews the same context the first one had |
| Audience split | The finished technical explanation | "Rewrite this for a customer with no technical background, in under 150 words." | One source of truth, several audiences, no cross-contamination |
| Clean summary thread | The answer with the final decision | "Summarize the decisions in this chat as a numbered list with owners and open questions." | The summary is based only on the conversation up to the decision |
Worked example: two launch announcements from one brief
- You spend six messages with a model agreeing on a product brief: audience, three key features, tone, and what not to promise.
- Under the answer that confirms the final brief, you select Fork from here. A new chat opens, titled "Fork of" plus your chat title.
- In the original chat you ask: "Write the launch email. Lead with the price change."
- In the fork you ask: "Write the launch email. Lead with the new feature, mention price only in the last line."
- You now have two drafts built on the same brief, and neither draft influenced the other. Paste the better opening into the other thread if you want a hybrid.
Without the fork, the second request would see the first email in its history and tend to echo its structure.
How Fork from here works in magicdoor.ai
| Behavior | What happens |
|---|---|
| Where it appears | Under assistant responses in your saved chats, once the response has finished |
| What is copied | Every message up to and including the response you chose; later messages are left out |
| New chat | Titled "Fork of" plus the original title, then opened automatically |
| Context kept | The original chat's skill and project |
| Files | Attachments on the copied messages are copied into the new chat. If a copy fails, the fork is still created and the confirmation tells you |
| Original chat | Unchanged |
| Cost | Creating the fork does not run a model. New messages in the fork are charged as normal usage |
Prompts to send right after forking
- "Before we continue, restate the goal and constraints of this conversation in three bullets." This checks the branch carries the context you think it does.
- "Ignore the last suggestion. Propose two alternatives that take a different approach."
- "Assume the plan above failed. What is the most likely reason, and what should we change?"
- "Continue, but write for [new audience]."
Limitations
- Fork is available in saved chats you own, not on shared chat views.
- The fork is a copy at that moment. Later changes in the original are not synced, and vice versa.
- A fork includes everything before the chosen response, so it does not remove earlier dead ends. Fork from an earlier response if you want them gone.
- Forking does not choose a model for you. The next message uses whichever model you select.
Try it on a chat you already have: open magicdoor.ai chat. See pricing for plans.
FAQ
How do I branch a conversation in magicdoor.ai?
Below any finished assistant response in a saved chat and select Fork from here. magicdoor.ai creates a new chat titled "Fork of" plus the original title, containing every message up to and including that response, and opens it.
Does forking change the original chat?
No. The original conversation stays as it was. The fork is a separate chat you can continue, rename or delete independently.
Are attachments copied into the fork?
magicdoor.ai copies files attached to the copied messages into the new chat. If a file cannot be copied, the fork is still created and the confirmation message tells you about the file problem.
Does the fork keep my skill and project?
Yes. The new chat keeps the original chat's skill and project, so the current instructions of that skill and project apply.
Does forking cost anything?
Creating the fork copies saved messages and does not run a model. You pay normal usage only when you send new messages in the fork.
Related Resources
AI Model Costs on magicdoor.ai: Current Chat Rates
Compare current chat-model input, cached-input and output prices, including Claude Haiku 5.5 long-context tiers and Mistral Large 4 guarded Gateway rates.
Mistral Large 4 (Le Chonk) on magicdoor.ai: Gateway Limits and Pricing
Compare Mistral Large 4 (Le Chonk)'s image and reasoning capabilities, Vercel Gateway context window, conditional privacy routing, and current token rates.
Getting Started with Skills
Create reusable instructions in a fresh conversation, review them, and use your skill in chat.
DeepSeek V4.1 Flash 1017 Guide - Context, Privacy, and Pricing
DeepSeek V4.1 Flash 1017 checkpoint guide: Fireworks USD pricing, optional GreenPT EUR pricing, 1,048,576-token context, tools, and provider privacy boundaries.