CONTEXT BUDGET

Long Claude chats get worse, not better. Why: context bloat. The signals to start a fresh chat, the trick to carry over what matters, and the 50k-token rule that pros follow.

THE PROBLEM

You start a chat with Claude. It's great. You keep going. By exchange 30, the answers are getting weaker. By exchange 50, you're frustrated.

You assume Claude got worse. It didn't. The chat got too long.

WHY LONG CHATS DEGRADE

Claude's context window is huge (1M tokens). But within a single conversation:

Earlier turns dilute later ones. Claude has to consider everything you've ever said in the chat, even tangents.
Bad outputs poison the chat. Once Claude went down a wrong reasoning path, it might keep referencing it.
Context drift accumulates. Small misunderstandings compound.

The cure: start a fresh chat. Bring forward only what matters.

THE 4 SIGNALS TO START FRESH

1. The chat is over 50k tokens. (Roughly 100+ message exchanges or 1 long-document analysis.) Performance degrades past this.

2. Outputs are getting WORSE, not better. If you're re-prompting more, with worse results, the chat is the problem.

3. You changed topic. Different topic = different chat. Don't ask Claude about meal planning in a chat about marketing.

4. Claude got something wrong and won't let it go. When you correct Claude and the next 3 outputs still reference the old wrong info, fresh chat.

THE TRANSITION RITUAL

Don't just close the chat. Carry over what matters:

Before I start a new chat, summarize for me:
- The 3 most important things we figured out in this chat
- Any decisions we made
- Open questions still unresolved
- Anything that would help future-me pick up where this left off

Format as something I can paste into a new chat.

Claude returns a summary. You start a new chat. Paste the summary as your first message. You've kept the value, dropped the bloat.

THE 50K-TOKEN RULE

Pros follow this rule:

Under 20k tokens → keep chatting
20k-50k tokens → start being selective; consider if you should split the topic
Over 50k tokens → start a new chat after the current request

How to check: most platforms show conversation length. If yours doesn't, count: ~1 page of text = ~500 tokens. 100 pages of conversation = 50k tokens.

WHAT TO DO INSTEAD OF SCROLLING UP

When you need to reference something Claude said earlier:

Bad: "Remember when you said X earlier?" (forces Claude to scan)

Good: Copy the specific quote and paste it back: "Earlier you said: 'X.' Build on that."

Saves tokens. Sharper output.

3 THINGS YOU CAN'T SKIP

1. Run the transition ritual EVERY TIME you start a new chat. It takes 30 seconds. Saves you from re-figuring out what you already figured out.

2. Use Memory (Day 26) for facts that should persist across all chats. Memory ≠ context. Facts you want Claude to ALWAYS know go in memory; situational context goes in the current chat.

3. Use Projects (Day 40) for long-running work. Same project, multiple chats, shared knowledge base. Better than one giant chat.

WHAT'S NEXT

Day 20 of 100. Tomorrow: Day 22 — The Hallucination Catcher.

SAFETY CHECK

Same as Day 1.