Conversation Architecture

Long conversations with AI get worse the longer they run. This reading shows you what that looks like, why it happens, and what to do about it.

Time to read: about 7 minutes

You may not have seen it yet. Most people run into it in the middle of their most important work. Recognizing it early is most of the skill.

Thirteen short pages, one idea each. Use Next below, or the arrow keys on a laptop.

Turn 3. It's sharp.

Here's a task you might hand to Claude. You're a VP of Business Development who needs an expansion proposal, and you start the right way, with a clear setup frame.

Claude honors every part of the brief. It stays under the limit, and it doesn't explain your product to people who already know it.

Turns 4 to 15. You make it better.

The first draft is a draft, so you push back where it misses.

Each round gets closer to what you actually need. Along the way you cut the limit to 1,500 words, try a financial appendix, and drop it.

Turn 22. It drifts.

An hour in, the proposal is in good shape. Now you ask for one more change.

1. It explains your product to a committee that already knows it. The audience instruction from turn 1 has faded.
2. It's back over your limit. The 1,500 words you set is gone too.
3. It cites the appendix you dropped.

None of this is dramatic, so it is easy to miss.

Turn 41. It agrees with everything.

You've spent twenty minutes correcting it. Each fix holds for a turn or two, then slips.

1. It agrees instead of engaging with what you said.
2. Its answers are getting shorter and more generic.
3. It's repeating itself.

Notice how you feel at this point. A spring participant named it the cuss-at-it indicator: when you catch yourself swearing at the tool, the conversation is already done.

Michele, spring 2026 cohort

The signs, in one place

• It repeats itself or circles back to earlier points.
• Its answers get shorter or more generic.
• You're correcting things it already knew.
• The output is drifting from where you started.
• It agrees with everything and stops pushing back.
• You're frustrated. If you're swearing at it, that's Michele's cuss test, and the conversation is done.

Any one of these is enough to act. First, know why it happens, because the cause tells you which fix will work.

The window fills up

Every conversation has a fixed amount of space, called the context window. Your messages, Claude's replies, and every document you share all take up room in it. As it fills, older material gets compressed or dropped to make space.

In the example, your word limit and your audience lived in turn 1, the oldest part of the conversation. They were the first things to blur.

The newest message wins

Even inside the window, Claude pays more attention to recent messages than to early ones. An instruction from an hour ago carries less weight than what you said a minute ago.

In the example, "Georgia goes first" was the freshest thing in the conversation, so Claude built the whole reply around it. Your standing instruction about the audience lost to your latest request.

Every wrong turn stays in the conversation

A long conversation keeps everything: the 2,000-word limit and the 1,500-word limit, the appendix you added and the moment you dropped it. Claude tries to reconcile all of it every time it answers.

Pushing harder does not work, because each correction is one more message to weigh against everything that came before.

"It's like trying to give somebody a suit that has previously been tailored, and at a certain point it's just too messed up and you gotta throw it out and start over."

Jeff, spring 2026 cohort

Correct once, specifically

When one thing slips, name it precisely and restate the instruction in the same message.

A specific correction often holds. It's also good practice: Anthropic's AI Fluency Index found that conversations with iteration and refinement were 5.6 times more likely to include users questioning Claude's reasoning, and 4 times more likely to catch missing context.

If it slips again, start fresh

Starting over feels like throwing away an hour of work. You keep the decisions and lose the clutter, as long as you carry the decisions forward. That's the next page.

Carry the decisions forward

Before you close the old conversation, ask Claude to summarize it.

The whole pattern

This week's homework asks you to let one conversation decay on purpose, then recover it. Now you'll know what you're looking at.

Claude's memory and Projects now carry some context between conversations for you. Two Memories, Not One covers what they keep and what they don't. Neither replaces this pattern. While this curriculum was being built, Claude's search pulled a conversation from November 2025 instead of the one from the day before. It used similar words but was a different piece of work, and a human caught the mistake. Treat what the platform remembers as a starting point, and correct it when it's wrong.