Editing Tokenization & Context Windows (section)

== Practical tips ==

'''Start fresh for new topics.''' Don't keep one mega-conversation running for everything. A new topic deserves a new conversation with focused context.

'''Front-load the important stuff.''' Put your most critical instructions, context, and constraints at the beginning of your prompt. Information at the start and end of the context window gets more "attention" than information buried in the middle.

'''Summarize before you continue.''' If a conversation is getting long and you want to keep going, ask the AI to summarize the key decisions and context so far, then start a new conversation with that summary.

'''Use persistent context features.''' Claude Projects, custom GPTs, and system prompts let you set context that persists across messages without eating into your per-message token budget. The [[The Handoff Protocol|Handoff Protocol]] exercise teaches you to design these.

'''Be selective with document uploads.''' Instead of uploading a 100-page document and asking a question, extract the 5 relevant pages. You'll get better answers and use less of your context budget.