@boiler

How To Save Claude Tokens Core principle: Claude re-reads the entire conversation before every reply. Longer threads = exponentially more tokens consumed. Prompt & Conversation Hygiene 1. Edit, don't follow up — Use the Edit button on your previous message instead of sendinga correction. Follow-ups stack; edits replace. 2. Batch your tasks — Send one message with three tasks instead of three separate messages. One context reload vs. three. 3. Start a new chat when the topic changes — Old messages are dead weight when you've moved on. 4. Summarize and reset every 15–20 messages — Ask Claude to summarize key decisions, copy it, and paste it into a fresh chat. 5. Restart from a prior message when things go sideways — Use "Restart conversation from here" as far back as possible rather than adding correction messages. 6. Ask Claude to ask you questions — A short prompt like "I want to [task]. Ask me questions before you start" costs far fewer tokens than a 500-word prompt repeated in every re-read. File & Context Management 7. Convert files before uploading — Extract text from PDFs/DOCX/PPTX and paste as plain text or .md . A PDF page = 1,500–3,000 tokens; clean text can be under 100. 8. Crop screenshots tightly — A full 1000×1000 image ≈ 1,300 tokens; a tight crop can drop it below 100. 9. Use Projects for recurring documents — Upload once, cache forever. Repeated uploads re-tokenize every time. 10. Keep system/context files lean — Profile or "about me" files should stay under 2,000 words. They're read every session. 11. Only load files relevant to the current task — Don't dump your entire folder "just in case." Every file read costs tokens. 12. Use a session-notes.md for continuity — At session end, ask Claude to write key decisions and next steps. Start the next session by reading that file. Model & Feature Selection 13. Match the model to the task — Haiku for quick questions, Sonnet for standard tasks, Opus for deep/complex work only. 14. Plan in Chat; build in the expensive tool — Do your thinking and iteration in Chat, then move to the file-creation tool once you know exactly what you want. 15. Turn off unused features — Web search, connectors, and extended thinking all add token overhead. Default everything off; enable per task. 16. Be specific with connector queries — "Search Slack from the last 7 days for Q2 launch messages" is far cheaper than an open-ended search. 17. Use /schedule for recurring tasks — Automate repeating reports/digests instead of running them manually in growing sessions. Output & Generation Control 18. Target edits, not full redos — "Only redo section 3, keep everything else" instead of "redo the report." 19. Suppress verbosity — Append "No commentary. No explanations. Just the output." when you know what you want. 20. Use structured preferences/styles — Set User Preferences and a response Style (e.g.,Concise) in Settings so you don't re-explain yourself every chat. Claude Code Specifics 21. Give Code a precise scope upfront — Code explores broadly if left open-ended. Specify the file, output format, and exact task before it starts. 22. Use CLAUDE.md for permanent context — Put conventions and recurring instructions there, but keep it short. Bloated CLAUDE.md gets ignored. Broader Strategy 23. Don't use Claude for tasks it's poor at — Image generation and real-time search burn tokens with poor results. Use purpose-built tools instead. 24. Spread sessions across the day — Claude's usage window is rolling (≈5 hours). Splitting morning/afternoon/evening sessions lets earlier usage roll off. Source: How to stop hitting Claude usage limits — Ruben Hassid, How to AI (Substack)

1 reply

0 recast

1 reaction