You hit the limit. Again.
It's 2pm, you're mid-flow on something, and Claude tells you you're out of credits. You either wait it out or reach for the credit card.
Here's the thing: most people don't have a usage problem. They have an efficiency problem. The same output is achievable for a fraction of the token cost if you change a few habits.
These are the ones that actually moved the needle for me.
1. Stop re-explaining yourself every session
If you're opening a new chat and typing a paragraph of context before every question, you're burning tokens before you've even started.
Fix: Use Projects. Write your context once in the Project instructions and Claude carries it into every conversation automatically. Your stack, your tone, your constraints — set and forget.
For Claude Code users, the equivalent is your CLAUDE.md file. Put your project rules there and Claude reads them on every session without you repeating yourself.
2. Be specific or be expensive
Vague prompts produce long, meandering responses that you then have to re-prompt to fix. That's two or three exchanges where one would have done it.
Compare:
- "Help me write an email" — Claude asks clarifying questions, writes something generic, you refine it twice
- "Write a two-paragraph follow-up email to a client who went quiet after a proposal. Tone: direct, not pushy. End with a single yes/no question." — done first time
Specificity is free. Vagueness costs tokens.
3. Use /clear in Claude Code (and mean it)
Every message you send includes all the previous context in that conversation. The longer a thread runs, the more tokens every new message costs — even if that earlier context is irrelevant.
When you switch tasks, run /clear. New task, clean slate, lower cost per message. It feels wrong to throw away context. Do it anyway.
4. Don't ask Claude to read files it doesn't need
In Claude Code, every file you add to context gets tokenised. If you're working on one component and Claude is reading your entire codebase, you're paying for context you don't need.
Be surgical. Reference specific files. Ask Claude to read only what's relevant to the current task. Read components/Button.tsx beats look at my whole project.
5. Batch your questions
If you have five small questions, don't ask them one at a time. Five separate conversations means five sets of context loading, five response preambles, five closing summaries.
One message with five numbered questions. One response. Done.
This is one of the habits that ADHD brains struggle with most — the impulse is to fire questions as they come. Slow down, batch them up, save the credits.
6. Use Claude for thinking, not transcription
One of the most common credit leaks: asking Claude to reformat, rewrite, or clean up text that's already basically right.
Transcription tasks (restructuring a doc, changing bullet points to prose, reformatting a table) eat tokens without adding much value. Use Claude for the hard thinking. Use a cheaper or faster tool for the mechanical stuff.
7. Write reusable prompts once, not every time
If you're typing the same prompt structure repeatedly — same context, same format request, just different content — you're wasting effort and tokens.
In Claude Code: Create custom slash commands for your repeated workflows. /draft-post, /review-pr, /summarise-meeting — write the full prompt once, invoke it with a word.
In Claude.ai: Save your go-to prompts. The Projects system and custom instructions exist exactly for this.
8. Use the right model for the task
Sonnet handles 90% of tasks at a fraction of Opus's cost. Most people default to the biggest model out of habit.
Use Opus when you need deep reasoning on genuinely complex problems. Use Sonnet for everything else. The quality difference on everyday tasks is negligible. The cost difference is not.
9. Stop asking Claude to do 10-second tasks
I have seen people ask Claude to write a regex to rename a file. The regex takes 10 seconds to Google. The Claude conversation takes 2 minutes and costs tokens.
Claude's value is in tasks that would take you 20 minutes to think through or 2 hours to write. If you can do it yourself in under a minute, just do it.
10. Set up skills so Claude knows what you want without asking
Every time Claude has to ask a clarifying question, that's a round-trip. Every time it produces something in the wrong format and you correct it, that's another round-trip.
Skills (Claude Code's built-in system for giving Claude persistent instructions and tools) eliminate this. You define how you want things done once. Claude follows it without prompting.
If you're building a system that works with your brain rather than against it, the ADHD AI Skill Pack is where I started. Twenty skills built around the way ADHD brains actually work — context management, task switching, the decision paralysis spiral. Worth it if you're burning through credits and still not finishing things.
The common thread across all of these: Claude rewards intentionality. The people who use it most efficiently aren't the ones sending fewer messages — they're the ones sending better ones.
Worth figuring out before your next top-up.