Table of Contents
First, the one thing that explains all of this
Claude doesn't count messages — it counts tokens, and it re-reads the whole chat on every single turn. A 30-message thread pays for all 30 messages again just to answer message 31. Almost every trick below comes down to one idea: stop making Claude re-read stuff it doesn't need.
The 12 tricks
1. Edit your prompt — don't reply "no, like this."
When Claude misses, your instinct is a follow-up correction. Every follow-up stacks onto the history and gets re-read forever. Instead, hover your original message, hit edit, and regenerate. The wrong answer gets replaced, not piled on. Over a long back-and-forth this is the single biggest saver.
2. Start fresh more often than feels natural.
A short, focused chat beats a sprawling one every time. When a thread has done its job, open a new one. You're not "losing progress" — you're dropping dead weight Claude was dragging into every reply.
3. Upload a file once — reuse it everywhere.
Re-attaching the same PDF across four chats means Claude re-processes it from scratch four times. Drop it into a Project's context once and every chat in that project reuses it. (Bonus move: paste heavy text into doc.new, download as Markdown, and feed the lightweight .md instead of a fat PDF.)
4. Right-size the model.
Running the heaviest model on everything is the quiet leak. Use the lightest model that can do the job — the fast one for quick answers and formatting, the mid one for everyday work, the heavy one only for genuinely hard reasoning. Most daily tasks don't need the flagship.
5. Dial down Effort + Thinking for simple tasks.
Higher effort and extended thinking spend reasoning tokens before Claude writes a word. Great for hard problems, pure waste on "rewrite this line." Turn them down (or off) by default; flip them up only when the task earns it.
6. Kill the toggles you're not using.
Web search and other tools inject hidden context into every prompt behind the scenes. Brainstorming? You don't need live search. Leave extras off unless the task actually calls for them.
7. Be specific: verb + scope.
"Make it better" forces Claude to guess what you mean and generate a menu of options you'll never use. Vague prompts are token multipliers. Name the exact action and the exact target:
Rewrite the second paragraph to be tighter — no preamble, just the rewrite.
8. Cap the output length yourself.
Claude defaults to thorough. If you don't need thorough, say so. Tack this onto any prompt where you know what you want:
Answer in 3 sentences max. No preamble, no explanation — just the output.
9. Batch related asks into one prompt.
Three separate chats = three full context reloads. One prompt with three tasks = one. Instead of "summarize this" → "now list the points" → "now write a headline," send:
Summarize this, list the 3 key points, and suggest one headline.
10. Let Claude interview you instead of typing walls of text.
Clicking options costs almost nothing; typing paragraphs of setup costs a lot. Push the context-gathering onto Claude:
I want to [task] so that [outcome]. Ask me any clarifying questions first, then do it.
11. Set your context once, not every chat.
Stop re-typing "I'm a [role], keep it concise, skip the basics" in every new thread. Put it in your settings/preferences once and it rides along automatically — thousands of setup tokens saved over a month.
12. Carry context forward with a notes file, not a re-paste.
For anything ongoing (a coding project, a long piece of writing), don't re-explain where you left off. At the end of a session:
Write a session-notes file with the key decisions, current state, and next steps.
Next session, start with: Read the session notes first. You keep continuity without re-loading the entire history.
Bonus: the Lean Prompt Starter Pack
Six reusable wrappers. Save them, swap the brackets, paste.
The brief (my default opener):
I want to [task] to achieve [outcome]. Constraints: [any]. Ask me anything unclear before you start, then deliver. No preamble.
The tight output:
[task]. One clean pass — no preamble, no options, no explanation. Just the result.
The scoped edit:
Change only [specific part] to be [specific change]. Leave everything else exactly as is.
Caveman mode (for quick throwaway stuff):
Answer caveman-style: no pleasantries, no filler, keep every fact and code block. [your question]
The batch:
Do all of these in one reply: (1) [x] (2) [y] (3) [z].
The hand-off:
Summarize this chat into a notes file: decisions made, current state, open questions, next steps.
The one-line takeaway
Every token you save is context Claude can spend on your actual work instead of re-reading noise. Lean prompts don't just save your limit — they get you sharper answers. Win-win.