A token ≈ ¾ of a word. Every new message makes Claude re-read the entire conversation — so the cost of each turn compounds fast.
98.5% of those tokens are just Claude re-reading history — every single turn.
As the context window fills, Claude gets worse at finding the right details — even though they're all still in the window. That's context rot.
The double penalty: worse recall → you re-ask and re-explain → even more tokens burned.
Most people work in the red. ↑
Ask Claude for a summary of everything decided so far
Open a new chat
Paste the summary back — continue from there
How do I check context usage in Claude.ai or Cowork?
You can't. No %, no meter, no indicator.
Don't try to monitor it. Build the habit of resetting early by default — any time the task shifts significantly, open a new chat.
Don't mix jobs in a single session — each one drags its history into the next. Research → Write → Review: one job per chat = small, sharp context.
Fast & cheap · simple tasks
Balanced · your default
Heavy reasoning · high-stakes only
Slash commands for power users:
/clearWipe the session
/compactSummarize & shrink the session
/reDrop a failed attempt
/btwSide question, off-thread
/contextCheck how full you are
CLAUDE.md <200 lines, max. Every line costs tokens — every session, forever.
The mistake: using Opus for everything — like hiring a senior consultant to answer your emails. Expensive and unnecessary.
Runs every step — drafting, researching, formatting, iterating. Constant, and cheap.
Never runs by default. Called only when stuck — hard reasoning or a judgment call. Weighs in, then the Worker continues.
A reference card — save it, pin it, keep it close.
Paste into ChatGPT or Gemini — it extracts everything it knows about you — import that context into Claude.