Stop adding context to your AI projects

Despite the insane (and slightly worrying) weekly advances of AI, the most common complaint I still hear is:

“ChatGPT doesn’t sound like me.”

If that's you, you've probably tried the most obvious thing:

Just add more context.

Style guides. Audience avatars. Old newsletters. Swipe files. Transcripts. Examples. More. More. More.

I remember being at dinner with a friend who showed me one of his AI projects in Claude.

He must have had 20+ documents inside it.

And one of my community students does something similar - collecting advice from the best YouTube strategists he can find…

…and putting all of it into one project.

It seems logical.

More knowledge = better responses, right?

Not necessarily, for a simple reason:

Eventually, your context starts fighting itself.

For staters, even your own style guide can contain contradictions:

• “Be concise… but use plenty of examples.”
• “Sound casual… but speak with authority.”
• “Don’t waste time with preamble… but set everything up clearly.”

Now imagine adding a newsletter swipe file from two YouTube strategists to your context pile.

One writes their hooks using "Target, transformation, stakes".

The other uses "Proof, promise, plan".

Who's right? Whose formula should ChatGPT or Claude use to write your hook?

And this is just the start of your troubles.

Imagine how these contradictions will build up once you start adding 15, 20, 30 documents to your projects.

The problem is, the AI doesn’t have a little judge inside it to decide which rule is correct.

• Sometimes it favours one.
• Sometimes it compromises.
• Sometimes the answer changes depending on your prompt.

And it only gets worse.

Because AI models become less reliable at finding and using the right information as the amount of context grows.

Large AI projects may pull in information they think is relevant, rather than actually reading every document in full.

That means the absolute gold dust you heard in a podcast and added to your scriptwriting project yesterday... might never get read by the AI.

Think of it like this:

Continually adding more context documents to an AI project is like adding endless ingredients to a recipe.

Each ingredient might be tasty. And some will work really well together.

But pretty soon, you'll end up with a big bowl of brown slop.

So... what's the solution?

Well, Anthropic (who makes Claude) explained this very helpfully:

Use the “smallest possible set of high-signal tokens.”

In other words:

1. Keep your context minimal.
2. Keep it relevant.
3. And every so often, clean it up.

So, if you've got a project in ChatGPT or Claude that's getting a bit out of hand, start a new chat and paste this:

“Review the context and instructions available in this project. Find duplicates, contradictions, outdated advice and low-value information. Recommend what to KEEP, MERGE, SHORTEN or DELETE. Then propose a smaller source of truth that preserves the useful information without the overlap.”

More context can help.

But only when it’s the right context.

That's all for this week.

As always, if you have any questions, feel free to reply.

Speak soon,
George 👋