Transfer From ChatGPT to Claude

Starting with AI - Interlude

Starting with ai post post3.5.

I started writing a five part article about working with Claude. I have talked about an overview, general chat, Claude chat projects, and Cowork. The next session should be Claude Code, but before I dive into that, I think it’s time to pause and dig into how to move your work from ChatGPT over to Claude.

I am sure there are a TON of ways to do this, however, these are the two techniques I use.

  1. Transcript port
  2. Artifact Handoff

Migrating from ChatGPT to Claude

When I first switched over to Claude I wanted to make sure I captured every single word and gave them to claude. So I used a web tool called MarkDownload. This tool takes a web page and turns it into a markdown text file and lets you download it to your computer.

I would then drag/open the file in Claude and have it read the conversation.

Now there is a downside to this method. And the downside is that this can chew up a lot of context. But what is context you ask?

What is Context?

The simple answer is working memory. When ChatGPT and other tools came out, they had very small contexts. They could handle a few hundred words and conversations that were not that long. Now the standard is almost a million tokens.

That’s cool you say, but what’s a token?

But What’s a Token?

A token is a word fragment. The historical genesis of most modern LLM’s (in part) is from believe it or not autocomplete. If I start typing a word, you want the machine to successfully predict what I will say next and suggest the word I mean, not the one I am mangling on my touch screen with my pickle fingers. AI engineers realized that these predictive systems could guess what part of a word could come next from a fragment of a word. They call these fragments tokens.

And, if they had more context in which that token was being used, they had a better chance of guessing what token would come next.

Consequently, context size as measured in tokens became the standard for how LLM’s can think. It turns out, it is important to keep an eye on context when you are working with LLM’s.

Why Should You Care?

The main reason you want to care about context and tokens is quality. It has been demonstrated that when an LLM context gets too full, the model begins to drift. It confabulates more, and forgets rules given to it earlier in the conversation. In other words, the LLM becomes forgetful.

There is a tool on github that gets under the hood of these tools and watches how well the LLM performs depending upon how full the context buffer is. As a result, they suggest, as of the time I am writing this article, that chat bot can be at its best when its context is below 30% full. About 50% you might want to start thinking about starting a new chat, or purging the context.

Why am I telling you that?

Because if you move over an entire conversation from ChatGPT to Claude… you will burn an enormous amount of context with unnecessary tokens.

Let’s face it. Most chats are heavy with unbalanced text exchanges. you talk a little, AI talks a lot. I noticed a pattern where the AI wrote 11 words for every one of mine. Not all of those walls of text have value. I would argue most don’t. This is why my dominant form of transferring knowledge between LLM’s is now what I call the markdown artifact handoff document.

Artifacts

I used three very specific words: Markdown. Artifact. and handoff. Why?

  1. Markdown - a readable text document that saves formatting intent. This is an extremely token efficient way to transmit meaning.
  2. Artifact - the idea of an artifact is that it is something decided. It represents your understanding at that moment in time.
  3. Handoff - the purpose of the document is to help another AI agent get up to speed fast.

Why use Artifacts? There are approximately 1.3 tokens per word.

Here is the real word count from a conversation I had with Claude this morning:

Words: 19,150 Characters: 118,050 Tokens (est): ~25,000

That is a CHUNK of text. Converted into an Artifact? Markdown Artifact for Handoff: 2,138 words or about 2,800 tokens. 1/9th the token space.

Something else you need to know about Artifacts

Using Artifacts is extremely powerful for “seeding” new conversations. A powerful design pattern I use, is to work with AI until we reach something like a “check point,” and then I stop and ask it to generate a markdown artifact for handoff. Which it does. I can do this multiple times throughout a conversation, but the point is to consolidate our understanding and make it transferable.

If you ask claude, How's your context He will answer. (or she if you think of Claude as Claudia, or it if you like a more Asimovian view of thinking machines). But the point is, paying attention to context can help keep hallucinations, conflations, and confabulations to a minimum. (who knew we would be using those words so much?)

Summary

Once you use chat, projects, and cowork in Claude you may want to start moving over some of your chats from ChatGPT. The best ways I have found to do that?

a. use a web clipper to grab an entire conversational thread. b. ask ChatGPT to generate a markdown artifact for handoff.

And if you don’t know markdown yet, I encourage you to start getting familiar with the concept, because you will want to know it when you start using claude code.

Example Prompt

Hey claude, what is a markdown artifact for handoff to another AI? What is an artifact?  What is a living artifact? And why do you do everything in markdown?

Until next time,

Scott