Unpopular Opinion: If AI Keeps Wasting Your Tokens, Get Better at Using AI

It seems like this sub has basically become a running list of: “OMG AI USED ALL MY TOKENS IN 12 HOURS DOING ONE TASK!!!” Okay... and? You asked it to do that task. Were you not looking at it at all during those 12 hours? Did you never check your usage and think, hmm... this seems like a lot, maybe something is going wrong here? And before this turns into a model war, this applies to every model out there. Claude, Codex, Gemini, whatever. I've used all of them. Garbage in, garbage out. I've asked different models to do some pretty complicated shit. One project was basically: deploy 45 servers on a Hyper-V farm, create a domain, join everything to the domain, and install/configure X, Y, and Z according to a pile of documentation. It ran for around 10 hours and used about 15% of my $200 plan . Then I asked it to build an HTPC environment from end to end. That ran for around 6 hours and used about 75% of my plan . Why the massive difference? Prompt, guardrails, good documentation, the tools I gave it, and agent seperation/management. Another example, I asked it to build a remote browser system where an RDS session displays the browser, but the actual rendering, processing and network traffic happen through a desktop WebView. That project was planned out in detail. I had specialized agents with small clearly defined jobs, and manager agents coordinating everything. There were roughly 3,000 lines of documentation and instructions telling the system exactly what it was building and how everything should work. That used around 35% of my $200 plan . Then one day I basically said: “You know what would be neat? Make me an app that manages all my Raspberry Pis from one interface. Give it all the features of ABC software.” And you'll never believe what happened... It burned through basically my entire usage in about 18 hours and still produced a shitty result. Shocking. Because I gave it shitty instructions. I barely planned anything out and basically told it "here's an idea, go build it." So yeah, it wandered around, made bad decisions, redid shit and wasted a ton of usage. If you give a person shitty instructions you're probably going to get a shitty result. If you then disappear for 12 hours, give them no feedback, no direction, no milestones, and don't tell them they're heading down the wrong path, the result is probably going to get shittier and shittier as the project goes on. Then when they finally deliver a pile of garbage, whose fault is that really? AI isn't magically exempt from this. And the constant: “It used ALL MY TOKENS in 12 hours!” Okay? I can probably burn mine in four. Congradulations I guess? How many sub-agents did you spin up? How many tool calls were being made? How much duplicate research happened? How many times did the context get compacted and lose something important? How many agents repeated work? How much time did it spend wandering down dead ends because you never gave it checkpoints or guardrails? “I used all my tokens” tells me almost nothing about how much useful work actually happened. Here's some perspective from that 45-server project. The client was billed $15,000 . The project was roughly: 6 hours of AI-assisted documentation/planning 10 hours of autonomous AI implementation 3 hours of actual human management 10 hours of AI-driven testing, troubleshooting, fixes and user issues Total actual human time invested was about 19 hours . AI cost was about $6 . That's roughly $780 of revenue per human hour invested , before accounting for the rest of the business costs. Six fucking dollars of AI usage helped leverage 19 hours of my time into a $15,000 project. If that AI usage had cost me $200 instead of $6 , I wouldn't have cared. If it had cost me $500 , I still wouldn't have cared. Hell, if the AI portion had cost $2,000 and still let me deliver that project with 19 hours of human labor, I'd have paid it and moved on with my life. Perspective. I'm not trying to win a contest for who can use the fewest tokens . I'm trying to make money and get work done. Thats the part that seems completely missing from alot of these complaints. The people who actually figured out how to use these systems effectively usually aren't posting: “HELP!!! IT USED 100% OF MY TOKENS!!!” They're saying, yeah I burned through my allowance, but it just did 30 hours worth of work for $200 and saved me or my company hundreds or thousands of dollars. That's a win. Who gives a shit that the token meter says 100%? And then there's: “But after all that work it STILL didn't fix the issue!” Yeah... and? If you hired a human for $200 and somehow got 30 hours of engineering work out of them, there is absolutely no guarantee they would have solved the problem either. And if the issue is incredibly complicated, the codebase is massive, the documentation is hundreds or thousands of pages long, and your prompt is basically a vague one-liner... what do you honestly expect? The useful metric isn't: “How many tokens did it use?” It's: “How much useful work did I get for the money?” And I already know what some of the responses to this are going to be: “Nope. Not me. I'm good at AI.” “I'm good at prompting.” “My prompts are great.” “The model is the problem.” Okay. If you're consistently burning through your entire allowance, getting garbage back, and then complaining that every model sucks... clearly you're not as good at this as you think you are. Sorry. Either your prompting isn't as good as you think, your project management isn't as good as you think, your expectations are completely unrealistic, or you have no idea how to do a basic cost/benefit analysis. Or probably some combination of all of them. And again, this isn't a Codex thing, or a Claude thing, or Gemini or whatever model you happen to be pissed off at this week. Garbage in, garbage out. Get better at breaking projects apart. Get better at documentation. Get better at giving agents narrow responsibilities. Get better at checkpoints. Get better at recognizing when an agent is going down a stupid path and stopping it before it spends six hours digging the hole deeper. And most importantly, get better at measuring value produced instead of staring at a token meter like it's the fucking gas gauge on your car. If your AI agent spent 18 hours running in circles because you gave it a vague one-paragraph prompt, no architecture, no documentation, no guardrails, no checkpoints and then walked away... Maybe the lesson isn't that AI usage limits suck. Maybe the lesson is that you suck at managing AI agents. And fortunately, thats something you can actually fix. Get better.