Ever since I started using AI, there's one word I've been hearing more and more.

Context.

At first, I wasn't really sure what it meant.

But as I started using tools like Claude Code, I kept hearing things like "Your context is getting too long," "You should compact your context," or "Maybe it's time to start a new session."

What does any of that actually mean?

The more I thought about it, the more I realized that I'd already experienced plenty of situations that might have something to do with context.

I just didn't know what to call them.

No Matter How Many Times I Asked, the Image Wouldn't Come Out Right!

One of my biggest frustrations with AI was image generation.

I'd start with a prompt and ask AI to create an image.

Sometimes the result would look pretty good!

But then I'd notice one little thing I wanted to change.

"Actually, could you just adjust this part?"

And then...

Wait. That's not what I meant!

The part I wanted to change stayed the same, while something I liked had changed instead.

So I'd ask again.

And again.

Sometimes, no matter how many times I explained what I wanted, the image just wouldn't turn out the way I'd imagined.

And image generation takes time!

You wait for the result, finally get to see it, and discover that something else has gone wrong.

So disappointing!

When I talked about this with people I knew, several of them said they'd experienced exactly the same thing.

Apparently, I wasn't the only one!

"Wouldn't It Be Faster to Start Over in a New Chat?"

At one point, I told someone that I'd heard it might be better to stop making endless revisions and start over in a fresh chat with a new prompt.

But that person wasn't so enthusiastic about the idea.

They'd already shared a lot of information and preferences with the AI.

After many conversations and adjustments, the AI was finally starting to generate images that reflected what they wanted.

Starting a new chat felt risky.

What if all that information was lost?

What if they had to explain everything all over again?

They wanted to stay in the same conversation.

And when I heard that, I thought, yeah, I can understand that.

They'd put a lot of time into getting the AI to understand what they wanted.

But if the revisions weren't working, wouldn't starting fresh be better?

Which approach actually makes more sense?

So, What Exactly Is Context?

In everyday English, context means the background information or circumstances that help us understand something.

When we talk about AI, a useful way to think about context is the information an AI model receives and can refer to when responding to us.

For example:

  • Previous messages in a conversation
  • Instructions and requirements we've provided
  • Documents or files shared with the AI
  • Information gathered during a task

Exactly what gets included depends on the AI tool or service we're using.

Context is what allows AI to follow a conversation without making us explain everything from the beginning each time.

Imagine we've been discussing an image, and I simply say, "Can you make it a little brighter?"

The AI may understand which image I'm talking about because of the conversation so far.

That's incredibly convenient!

Context isn't something bad that we should always try to eliminate.

It's actually one of the things that makes working with AI so useful.

But AI Can't Handle Unlimited Information at Once

This is where another term comes in.

Context window.

A context window is the amount of information an AI model can work with at one time.

AI doesn't necessarily have unlimited access to every word we've ever said.

There are limits, and those limits vary depending on the model and service.

As a conversation gets longer, more and more information can accumulate.

The original instructions.

Requirements we changed halfway through.

Ideas we tried.

Attempts that didn't work.

Old decisions that no longer apply.

Of course, a long context isn't automatically a problem.

Sometimes knowing the entire history helps AI make better decisions.

But sometimes, that history can become complicated.

What happens if an instruction from the beginning conflicts with something we said much later?

Which one should the AI follow?

The more information we add, the more important it becomes to keep that information clear and relevant.

So maybe giving AI more and more context isn't always the answer.

Sometimes My Long Chats Became Incredibly Slow

There's another problem I've experienced many times.

Sometimes, after working in the same chat or session for a long time, AI responses started taking much longer.

I feel like this happens less often nowadays, but there were times when I had to wait an incredibly long time for an answer.

Eventually, I'd ask for a summary of everything important we'd discussed.

Then I'd take that summary into a new chat and continue working there.

And suddenly...

So fast!

The responses felt dramatically quicker.

I remember thinking, "Why didn't I do this sooner?!"

One possible reason is that longer conversations can require an AI model to process more information, which may increase response time.

But context isn't the only factor. The model, the tools being used, and the service's current workload can all affect speed.

Still, in my own experience, moving to a new chat with a summary made a noticeable difference on multiple occasions.

What Does /compact Do? Is It Different From Starting a New Chat?

After I started using Claude Code, I began hearing about a command called /compact.

It's a way to summarize a long conversation and free up space in the context window.

Instead of keeping every detail of the conversation exactly as it happened, Claude Code creates a summary of the important information.

That way, you can keep working in the same session.

It's a little different from starting a completely new conversation.

Here are a few options available in Claude Code:

Method What it does
/compact Summarizes the conversation so you can continue with less context
/clear Clears the current conversation context so you can start a new task
New chat or session Starts fresh, with any necessary information provided again
/btw Lets you ask a quick side question without adding it to the main conversation history

I'd also heard about /btw.

It stands for by the way.

I love how literal that is!

In Claude Code, it lets you ask a quick question related to your work without adding that exchange to the main conversation history.

Looking at these features, I started to realize that managing context is an important part of working with AI.

But there's something to watch out for.

When a conversation is compacted into a summary, some smaller details may be left out.

If you have important requirements, decisions, or rules, keeping them in a separate file can be a good idea.

If You Start a New Chat, Does Everything You've Shared Disappear?

This brings me back to the person who didn't want to leave their image-generation chat.

They were worried about losing all the information they'd shared.

And that concern makes sense.

A new chat doesn't necessarily inherit everything from the previous one.

But starting fresh doesn't have to mean throwing everything away.

For example, you could ask the AI in your original chat:

"Please summarize the character's confirmed features, the design requirements, and everything that must not change. Leave out rejected ideas and unsuccessful attempts."

Then you could take that summary, along with the reference image you want to preserve, into a new chat.

That way, you're not starting from absolutely nothing.

You're carrying forward the information that matters.

Of course, a summary can't guarantee that every detail will be preserved.

And starting a new chat won't magically fix every image-editing problem.

But there is a third option between keeping absolutely everything and throwing everything away.

Keep the important information, and start fresh with a clearer set of instructions.

Is Every Failed Image Edit Really a Context Problem?

Here's an important distinction.

Imagine asking an image-generation AI to change only the background while keeping a character's face exactly the same.

But the face changes too!

That doesn't necessarily mean the conversation has become too long.

Image-generation models can struggle to change one part of an image while preserving everything else precisely.

This can happen even on the very first revision.

On the other hand, if a conversation contains lots of old instructions, conflicting requests, and abandoned ideas, organizing that information might help.

The two problems can look similar from the user's perspective.

You keep asking for revisions, and they don't work.

But the underlying reasons may be different.

Understanding that distinction can help us decide what to try next instead of repeating the same request over and over.

Context Isn't Simply Better When It's Longer or Shorter

The more I think about it, the more interesting this becomes.

I understand why someone wouldn't want to lose all the information they've spent so much time sharing with AI.

But I also know how frustrating it can be when a long conversation stops going smoothly.

Both sides make sense.

Maybe the real goal isn't to make context as short as possible.

It's to keep the information that matters and organize or leave behind the information that no longer helps.

These days, I'm spending more time building apps and working on different projects with AI.

And I'm realizing that it's not just about what I tell AI.

It's also about what I preserve, what I update, and what I decide no longer needs to be part of the conversation.

One thing in particular has changed the way I think about starting fresh.

Starting a new chat doesn't necessarily mean throwing away all the work you've done.

You can keep your decisions, your files, your finished work, and the experience you've gained.

Sometimes, all you need is a fresh conversation.

"Context" sounded like such a technical word when I first heard it.

But maybe it's something those of us using AI have been dealing with all along, even before we knew what it was called!