How to never hit your AI usage limits

One of the most common complaints I see about Claude and ChatGPT is people hitting their limits.

I’ve been using AI agents daily for the last 12 months and I don’t think I’ve hit a limit once. Not because I use them less than everyone else. I use them a lot. It’s because without really thinking about it I’d picked up a handful of habits that cut down how much I burn through.

Here’s what they are, and why they work.

First, understand what you’re actually paying for

This is the bit most people don’t know, and everything else follows from it.

You’re charged per message. Not per conversation. And every time you send a message, the whole conversation gets sent along with it.

So message one is cheap. Message thirty is expensive, because it’s carrying everything that came before it. The longer a chat goes on, the more each new message costs you. It compounds.

Same goes the other way. The longer the answers you get back, the more you use.

Once that clicks, most of the tips below become obvious.

Start new chats more often

Long running conversations are the biggest drain there is.

If you’re asking for something that doesn’t need any of what you talked about earlier, start a new chat. You’re not losing anything, and you’re sending one message instead of one message plus forty old ones.

I know people like keeping one big thread going. It feels tidy. It’s also the single most expensive way to use these tools.

Change the model

I thought everyone knew this one. Turns out a lot of people don’t know you can even change the model.

You don’t need the smartest, most expensive model to look something up or rewrite an email. Use a lower one for everyday stuff and save the top model for things that genuinely need thinking. The cheaper models are much lighter on usage and they’re perfectly good at ordinary jobs.

Avoid dropping in full PDFs

PDFs eat tokens. Way more than people expect, because they assume it’s just text in there.

If you only need one section, copy and paste that section in. Use the bit you actually need as context instead of handing over the whole document. Only upload the full thing when you really have to.

Be careful with connectors

Connectors and MCPs can be heavy. The agent ends up holding a lot of context it doesn’t need.

This one varies. Some are light, some are enormous. But if you’ve got several running in one conversation it adds up fast. Turn off the ones you’re not using.

Tell it exactly where to look

When you’re working with files on your computer, point the agent at the actual folder. Better still, name the file.

You almost certainly know the file name, it’s right there in your file browser. Give it to the agent and it goes straight to it, instead of reading through everything you own trying to work out which one you meant.

Put everything in the first prompt

This is the one that makes the biggest difference for most people.

Don’t say “I want to build a photo sharing app”. Say “I want to build a photo sharing app with a home feed, a popular feed and user profiles. Users can upload one photo per post and add a caption.” Keep going. The more you put in, the better the chance it gets it right first time.

Every round of “actually, can you also” is another full conversation sent back and forth. Some back and forth is normal when you’re building something. But every bit you can avoid by being specific up front is usage you don’t spend.

Use sub-agents

This is the big one for me, because I’m using coding agents most of the time.

What I do now is use the top model as the planner and let it hand the actual work to sub-agents running on cheaper models. The smart model does the thinking, the cheaper ones do the typing.

It sounds complicated. It isn’t. You don’t touch any settings. You just add a couple of lines to the end of your prompt:

Use Fable as the planner and orchestrator. Create multiple sub-agents to complete the task and decide which tasks get assigned to which sub-agents. Use lower models to save on tokens wherever possible.

That’s it. It reads that, plans the architecture itself, then spins up sub-agents in phases to do the build. You can click in and watch them running.

This is how I’ve built pretty much everything over the last few months. Most of what I’m making doesn’t need a frontier model doing every single piece of it. I want the smartest model planning, because that’s where being smart matters, and everything else can run cheaper.

Upgrade if you can afford it

I’m on the $200 a month plan for both Claude and ChatGPT. Two subscriptions.

If you run a business, that’s a bargain for what you get back. I’m not going to pretend it’s nothing for everyone, but measured against what it replaces it isn’t a close call for me.

What this has let me build

None of this is theory. Here’s what came out of it.

Link Lowdown is a database and directory of interesting things across the web, with user submissions, automations and a leaderboard companies can pay to appear on. It handles payments, accounts, invoices, all of it. Built with Claude. A few years ago that’s weeks of work, maybe months.

aimanual.app is newer. It’s a hub for AI guides, teaching what me and my team have worked out, with members only content. Still not completely finished but there’s a lot on there already.

Then PlatSupply, my digital product store. It’s on Shopify, but AI runs almost all of it. The site, the automations, product delivery, customer service, design, email campaigns. I’ve automated as much as I can so it makes money without needing me in it every day.

Plenty of other things too. And I still haven’t hit a limit.

I don’t think that’s luck. It’s just these habits, most of which take no effort once you’ve done them a few times.