Stop Nickel-and-Diming Your AI Tools and Go All In

The Hail Mary

Things were looking rough with one of my clients. I was gearing up to fire myself and pivot to a variety of other red hot leads for exciting things. I didn't want to do that. I really hate these kinds of failures. But I've gained the wisdom through experience to know that sometimes the best thing you can do is accept a form of failure, drown the negativity out in positive things going forward, and just dwarf anything negative. If you get bogged down trying to fix a situation that's spiraling out of control, you usually make it worse and choke out all the possible positive things that could help heal all around.

I'm not going to get into the details of the back and forth. I've got my opinion, he's got his. I understand and I'm sympathetic to his perspective, but I stand firm on mine. There was crazy scope creep and the whole project plan got thrown out the window early on, which matters. But I just wanted to find a way where we could both move forward.

So I made an interesting deal. I needed to buy myself some time to work forward some leads that were looking good. And I had a suspicion, based on what I'd heard other people say and the experience I'd been having, that there was at least enough of a difference between the Pro Max subscription and what I'd been using. I told the client: look, I won't ask you for any more money. Gift me a $200 Claude Code Pro Max subscription and we'll finish this thing up. Your concerns with overages are addressed, you don't feel like you're getting screwed, and we move on.

The Nickel-and-Dime Setup

To explain why that deal even occurred to me, you need to understand what my setup looked like before. I had small subscriptions to a variety of AI tools. $20 to GPT. Something like $20 to Claude. About $10 in Grok spend. A Cursor subscription that had crept from $20 to about $60 with overspend allowance. And I'd set up a local LLM with Ollama on my new M4 Mac running the Qwen model, because I kept hitting limits and needed something to handle basic tasks.

I was trying to take these low-paid subscriptions across different models and spread them all out. I'd use GPT for high-level engineering discussions about what to build. Then I'd flip to Cursor to analyze the codebase and ingest the GPT markdown, because Cursor has a pretty good code-based indexing system. It would update the engineering design document to factor in the actual codebase. Then I'd flip over to Claude Code to plan and build whole features. Then back to Cursor for debugging and revisions, since that was the tool I was most familiar with.

And I was still hitting limits. I was authenticating with the AWS CLI, the Vercel CLI, the Supabase CLI, the GitHub CLI, just telling it in plain text what to do, and burning through tokens. So I set up the local LLM to offload some of that. The Qwen model is actually pretty decent in the context of a local LLM running on a decent Mac.

I'm not saying all this because it's instructional. I'm saying it because the real point of this whole talk is that you should not do that.

Throw It All Away

If you're where I was and you're trying to take flight with these AI-assisted development tools, these coding agents, subagents, and specialties, then throw out all of your little subscriptions and don't waste any time on a local LLM. Do whatever it takes to scratch out $200 and get yourself a Claude Pro Max. Or if you don't like Anthropic for some crazy reason, get the $100-200 Codex plan.

You are not going to use all of your quota until you crack automating subagents to the point where they're working 24/7. Everyone needs to know this. The limits are not clearly documented. They're difficult to understand. I still don't fully understand them. You have to push them yourself and see what you can get them to do. And what you're going to find is that in order to have a chance at using all of your quota through Claude Code, you're going to have to crack automated subagents that run on their own.

What Actually Happened

My math said that even if Pro Max was no more powerful than what I'd been using, it would at least give me a high enough token limit to finish the project and catch up on my other work. But what ended up happening changed my life.

It's a combination of things. Getting that Pro Max subscription and driving it around, realizing it's way faster, way smarter, and the token limit is so high you wouldn't believe it. I can't use it all myself. I've had to automate a personal assistant in order to do it. I've had to open up multiple terminals simultaneously working on multiple projects, each one spinning up its own subagents depending on what I've told it to do. And I'm still not on the verge of using it all.

Around the same time, OpenClaw blew up. I set it up and got a taste of what it feels like to have a heartbeat behind your agent and a good memory system. OpenClaw forced Anthropic to ramp up and improve their skill system, their memory system, and add scheduled tasks. A lot is happening really fast. But between this desperate situation that caused me to think of this crackpot scheme and being red-pilled on how good Pro Max actually is, that was a forcing function for me to try to automate subagents and build proactive agency.

I'd avoided OpenClaw because of the obvious security risks, but I'd recently bought a new Mac, so I figured I could keep everything sandboxed. What I found is that the heartbeat is the big thing. Proactivity changes everything. So I cloned the OpenClaw repository and built my own version, because Anthropic had shut down Pro Max authentication through OpenClaw, funneling everybody back into Claude Code. But I didn't want to wait for Anthropic to build all these features. I used my previous OpenClaw config as context and built my own version so I could keep using my Claude Code subscription with proactive agency.

What Gus Does Now

My agent's name is Gus. He's got his own Google account, his own Gmail. I wake up to a daily SEO audit report delivered via Telegram with links to GitHub pull requests. There's a weekly one that goes deeper because you don't want to spam certain checks every day. Gus sends me an email with a full breakdown and overview. The SEO audit agent codes out its own recommendations and opens PRs for me to review. This all happens while I'm asleep.

I built a bot that checks Upwork job postings, saves them based on filters from conversations we've had about our business strategy and my professional background. It self-heals the filters based on context. It gathers postings I might be interested in and sends me a Telegram message with the report. I've been talking about doing that forever, and now it's just running. Because all of this is in memory, I can go back and have conversations about what skills I should front on my Upwork profile based on job posting trends.

I added a web search skill, an email skill, a whole module of Upwork skills, a whole module of SEO skills. While I'm driving around, every time I have an idea, I pull out my phone, open Telegram, hit the voice-to-text button, and just start talking. I tell it all my ideas and hit send. It goes back and looks through all of our memory, everything it has insight into across our entire organization and operations, does web searches, and goes to work.

Today marks day 7 of my Claude Code Pro Max subscription.

Just Do It

If you feel like you are on a sinking ship as a developer, and you're afraid that you don't know what to do, and you know there's something powerful that a lot of people know about that you don't, and no one's crystallized it into words yet, I'm not here to give you those words. I don't feel like I can articulate it for you. I think it's different for everybody.

But what I will tell you is this: take a week. Take a day. Go all in. Scrape out $100, preferably $200. Buy whatever the highest pro plan is. I used Claude, I'm going to recommend Claude. Go clone OpenClaw and really push Claude Code to do as much as it can. Be open with it. Everything you've thought about, everything you've dreamed about. Have an open and honest conversation with this thing, almost like a financial psychiatrist. You are not going to look back.

Everyone has their own unique red pill moment with this technology. People say "you can just do things" and it's not going to make sense until it makes sense. You're going to have your own epiphany. You're going to have your own aha moment. It changes everything and you're going to be all right.

Throw out all those little subscriptions and go all in on one. Don't bullshit around. Don't nickel and dime it. Don't try to be cheap and conservative. Go all in, and you're going to find that in order to use the whole subscription, you're going to have to automate subagents. That is the key. It's a forcing function to make you think: how do I automate everything?

I haven't had a breakthrough like this in probably 10 years, since I left manual labor and trade work and went all in on coding and freelancing. This is a moment on par with that. I feel alive in a way that I haven't felt in a long time.

One last thing. I think what Anthropic is trying to do with their limits is brilliant. They're trying to shame you into automating things, but then constrain you into red-pilling other people. You're going to have to learn automation in order to use all of your subscription. But once you learn automation, you're going to want more subscriptions. And I think they're also capping you before you take humans out of the loop altogether. It's enough to force you into learning automation, but not enough to replace people entirely. It's very Anthropic. And if I'm right about how I interpret that, it's actually quite brilliant.