26/06/2026
Stop Hitting your Claude limits.
I have seen many posts about high token usage and hitting their limits. That’s why I thought of making this post that’ll help you avoid hitting your Claude limits.
Here are 15 mistakes I see every day — and the exact fix for each one.
───
Mistake 1 — Using Opus for everything
Opus is the most capable model. It’s also the most expensive on your token budget.
90% of tasks can be handled at the same quality level by cheaper models.
The right way to think about it:
Haiku 4.5 — quick lookups, formatting, anything repetitive Sonnet 4.6 — everyday writing, summaries, most daily work Opus 4.8 — complex reasoning, long documents, decisions that matter
You don’t need a sports car to go get groceries. Wrong model = wasted tokens before you’ve done anything useful.
───
Mistake 2 — Leaving connectors on by default
Connectors are powerful. Left on permanently, they’re expensive.
Every connector reads your tools the moment a chat starts. Leave Gmail, Slack, and Drive all on by default and Claude is processing thousands of tokens before you’ve typed a single word.
Fix: turn everything off by default. Turn connectors on per conversation, for the specific task that needs them. One conversation, one purpose.
───
Mistake 3 — Writing 500-word prompts
More words in a prompt does not mean better output. It means more tokens burned and more room for Claude to get confused.
29 words is enough for most tasks:
“I want to [task] so that [goal]. Read my files first. Ask me questions before you start.”
That last sentence changes everything. Claude stops guessing. You stop fixing outputs that missed the point.
───
Mistake 4 — Saying “redo the whole thing” to fix one part
You got a 600-word output. Section 3 is wrong. You type “redo it.”
Claude rewrites everything. Burns twice the tokens. You lose the parts that were fine.
Fix: be surgical.
“Only rewrite section 3. Keep everything else exactly as it is. No commentary. Just the output.”
───
Mistake 5 — Sending 3 messages for 3 tasks
Every message forces Claude to reprocess the entire conversation again. Sending multiple follow-ups is one of the biggest hidden token drains.
Fix: batch your tasks into one message.
“Do three things: summarise this document, list the 5 key points, and suggest a headline.”
One message. Three outputs. A fraction of the token cost.
───
Mistake 6 — Stacking corrections on top of each other
Claude does something wrong. You type “No, I meant…” It’s still wrong. You type “Actually what I wanted was…” Now your context is polluted with failed attempts and Claude is trying to reconcile all of them.
Fix: click Edit on your original message. Fix the prompt. Regenerate. The bad version disappears from the history entirely. You’re not adding a correction — you’re replacing the original.
───
Mistake 7 — Letting chat sessions run too long
LLM performance degrades as the context window fills. When it gets full, Claude may start forgetting earlier instructions or making more mistakes. Anthropic
Most people notice Claude getting worse in a long conversation and think it’s a model problem. It’s a context problem.
Fix: every 15-20 messages, ask Claude to summarise the key decisions and context into a short brief. Copy it. Start a fresh session. Paste the brief at the top.
Same quality from message 1. Every time.
───
Mistake 8 — Uploading PDFs directly
A raw PDF is one of the most token-heavy things you can feed Claude. Images embedded in PDFs, metadata, formatting characters — all of it counts.
Fix: copy the text out of the PDF and paste it directly. Or open the PDF in Google Docs, download as plain text or markdown, and paste that instead. Same information. A fraction of the tokens.
───
Mistake 9 — Dumping all your files into every session
More files does not mean better context. It means Claude is reading things that have nothing to do with your current task.
Fix: only include what this specific task actually needs. Precise prompts, focused sessions, and minimal file loads — every narrowing reduces waste and improves output quality.
For a quick email draft, you need zero folders. For a client proposal, you need the client brief and your template. Nothing else.
───
Mistake 10 — Keeping multiple topics in one chat
Three unrelated questions in one chat means Claude is carrying all three topics in context for every single reply.
Dead context is dead tokens. New topic means new chat. Always.
One conversation. One purpose. When the topic shifts, open a new chat. Your token budget will last significantly longer.
───
Mistake 11 — Building your about-me file without limits
Your about-me file is supposed to give Claude context about who you are. A 10,000-word about-me file gives Claude too much to process and too little to prioritise.
Fix: keep it under 2,000 words. Every line should change how Claude responds to you. If a line doesn’t do that — cut it. Context files are tools, not diaries.
───
Mistake 12 — Moving to Cowork before you know what you want
Cowork is powerful for ex*****on. It’s expensive for exploration.
If you don’t know exactly what you want yet, plan in Chat first. Think through the task. Refine what you’re asking for. Then move to Cowork once you have a clear brief.
Planning in Chat costs almost nothing. Rebuilding a Cowork output three times costs a lot.
───
Mistake 13 — Uploading the same file to multiple chats
You have a client document. You paste it into Monday’s chat. You paste it again into Tuesday’s chat. Again on Wednesday.
Every paste is the same tokens burned again.
Fix: use Projects. Upload the file once to the Project. Every conversation inside that project references it automatically — without re-uploading it every time.
───
Mistake 14 — Never setting Global Instructions
Without Global Instructions, you repeat the same preferences in every single conversation.
“Keep it short.” “Write in plain English.” “Don’t use bullet points.”
Every chat. Every time.
Fix: Settings → Global Instructions. Write it once. Every conversation follows those rules permanently — no setup required, no reminders needed.
───
Mistake 15 — Using Claude for things it can’t do well
Claude is not the right tool for everything.
No image generation — use ChatGPT. Real-time social media data — use Grok. Quick voice-to-text — use a dedicated dictation tool.
Using the wrong tool doesn’t just give you a worse output. It burns your tokens on a task Claude was never designed for. Know your tools. Open the right one. Get the result faster.
───
You don’t need a bigger plan.
You need better habits with the one you have.
Fix the top 3 on this list this week. Your token budget will feel completely different.
───
Which of these are you doing right now?
