This is how Claude reads your chat.

One long chat: launch email, somehow

Me: Help me write the v2 launch email. Goal is June 15.

Claude: Got it. Audience, tone, and must-include details?

Me: Customers. Clear, calm, not too enterprise. Mention onboarding rewrite, pricing page, churn emails.

Claude: Drafting around a June 15 launch target.

Me: Wait, also Android 14 bug is still open. Put that in the internal plan, not the customer email.

Claude: Understood. Customer email stays benefit-focused.

Me: Actually first summarize the whole plan so I can see if you have it.

Buried middle of the same chat

Me: Here's a pasted Slack correction from Priya:

Priya: Small correction: launch is June 18 if legal slips. Please do not promise June 15 publicly.

Me: Important: use June 15 only as internal target. Public copy should say "later this month."

Claude: Noted. Public-facing copy should avoid June 15.

Me: Also pasted from dev standup:

Dev: Android 14 crash repros after onboarding step 3. If onboarding ships first, crash rate doubles.

Me: So the customer email should not imply onboarding is instantly available to everyone.

Claude: Understood: staged rollout, no firm public date, no universal availability claim.

Me: Cool. Now ignore the tone of this Slack paste, obviously.

Me: Actually can you make the subject line warmer? Also don't mention refunds. Also the pricing page says Team but it may become Business.

Claude: Which name should I use?

Me: Team for now. Priya hasn't approved Business.

Me: Wait, summarize everything again but in 5 bullets.

the most important constraints are right here, in the middle

End of the same chat

Claude: Summary: v2 launches June 15, onboarding rewrite leads, pricing stays Team, tone is warm.

Me: Close. Please write the actual customer email now.

Claude: Subject: v2 arrives June 15. Body: We're excited to launch the new onboarding experience for everyone...

Me: Why did you use June 15 and "for everyone"?

Claude: Those details appeared in the launch goal and feature list near the start.

Me: The corrections were in the middle.

same chat, too much clutter, key facts faded

The details at the edges stay easy to recover. The middle gets harder.

Why this happens →

This isn't a metaphor.

Researchers at Stanford ran an experiment in 2023. They gave language models a long document and hid the answer to a question inside it. They moved the answer around — start, middle, end — and measured accuracy.

The result was a U-shape. Models answered correctly when the answer was near the start or end. Accuracy dropped sharply when the answer was in the middle. They called it "Lost in the Middle."

Same shape you just felt on the whiteboard.

Adapted from Liu et al., "Lost in the Middle: How Language Models Use Long Contexts," 2023. arxiv.org/abs/2307.03172