How to Reduce Your Claude AI Token Usage: 11 Proven Habits

The one-line mental model

Higher token usage simply means more effort — Claude doing more work, or processing more information. That's either more output, or more of the conversation history getting re-sent on every single turn, since the underlying API is stateless per request and history compounds as a chat gets longer.

Six habits to start with today

• Put obvious, standing facts in system instructions once — not re-explained every chat.
• Turn on Reflect so Claude already carries context forward, instead of you re-typing background each time.
• Type or paste directly / import from other AI chats rather than re-describing something from scratch.
• Review and edit before sending — a clean, correct prompt avoids a round of back-and-forth correction turns, each of which resends the whole thread.
• Keep a short 'who's who' in memory or system instructions instead of re-establishing that in every chat.
• Avoid PDF, DOCX, PPTX, XLSX. Prefer MD, JSON, HTML — these formats tokenize far more densely and cleanly.

Five more — the high-leverage ones

These five matter more than the first six combined for anyone doing real, repeated work with Claude:

• Don't ask for an Artifact unless it needs to be interactive — for a mind map or flowchart, ask for a plain HTML file instead. Artifacts carry real overhead because they need complete, self-contained, runnable code every time.
• Avoid long-running chats in one window — more history means more gets resent every turn. Start a new chat per task.
• Convert PDF and DOCX to Markdown before uploading — can get you close to a 90% token reduction on that document.
• In Cowork: paste the file names and ask for a batch/command-prompt script to run yourself, rather than having it operate on the files end-to-end.
• Ask for a Python script and self-run it, instead of having Claude execute the whole task itself — the single highest-leverage habit on this list, realistically 80-90% token savings on repetitive or file-heavy work, because you pay once for the code instead of once per tool round-trip.

Why 'self-run instead of agentic' works so well

Every time Claude executes a step of a task itself — reading a file, running a check, writing an output, reading the result back — that round-trip costs tokens. A Python script does the same mechanical work without any of those round-trips. You pay Claude once, to write the code; the code then runs for free, as many times as you need. This is the exact principle behind the Excel Dashboard Pattern and PDF Summary Pattern in our data-security guide, which double as security controls for the same reason.