Skip to content
Claude and ChatGPT’s latest models are nitpicky and burning tokens. Here’s the fix

Claude and ChatGPT’s latest models are nitpicky and burning tokens. Here’s the fix

I’ve been using two of the newest Claude and ChatGPT models to tinker with my personal self-hosted projects… and lately, I’ve found myself drowning in bugs.

Not that Claude Opus 5 and GPT-5.6 Sol (which powers the new GPT Work agent) are causing bugs in my code. Rather, they’re finding them… everywhere they look. Indeed, my queue of open “issues” seems to have doubled in the past week, filled with critical flaws and elaborate workarounds, each broken down into multiple steps.

Even worse, when I ask one of the models to review the issue of another, the reviewing model invariably finds all kinds of problems with the first model’s work, leading to yet more multi-step fixes.

All these fixes are taking a toll on my usage limits. I’m a Claude Pro and ChatGPT Plus subscriber ($20/month each), and lately I’m blowing half my weekly allowance chasing bugs found by Claude Opus 5 and GPT-5.6.

I’m all for being thorough, but the nitpicky nature of the new Claude and GPT models is akin to a body shop mechanic who wants to replace the hood of your car after spotting a tiny nick in the paint. And this isn’t just an issue for coders, by the way—it’s a problem for anyone who needs help with editing work proposals, reviewing Excel spreadsheets, or performing any number of everyday AI duties while on a token budget.

So I asked Claude Opus 5 for an “on a budget” prompt that takes a task and asks the AI to respect your usage limits, separating real problems from cosmetic ones and scaling down the scope to something that’s considerate of your time and wallet.

Naturally, Opus 5 spat out a fairly lengthy six-point plan for the prompt, and when I asked GPT-5.6 Sol to review it, it gave me a dozen detailed criticisms. Telling GPT-5.6 to go back and follow the instructions in the original prompt, it quickly backtracked (“You’re right—I gold-plated the review”) and narrowed the list to three.

Here’s the final prompt that Opus 5 and GPT-5.6 Sol agreed upon, pared down to four bullet points from the original six (and yes, it borrowed the “body shop” metaphor I originally gave it):

Before you start, treat my time and usage budget as a real constraint. Match the effort to what the task is actually worth — like a good mechanic, not a body shop upselling a new hood over a paint chip.

– **Smallest fix that fully solves the stated problem.** No scope creep, no “while we’re here” refactors, no gold-plating. If you’re about to propose something elaborate, first check whether a 3-line version gets 90% of the value.

– **Separate what matters from what doesn’t.** Say what’s worth doing now, what’s minor, and what to leave alone. Recommending less is a win; don’t pad the list to seem thorough.

– **Rank by value-for-cost.** Give a rough effort (S/M/L) per item, put cheap high-impact things first, and flag anything whose fix costs more than the problem it prevents.

– **Verify cheaply.** Check the claims that would change the plan; don’t audit everything.

Unless I explicitly ask you to execute, give me a short ranked plan and stop.

I tried this prompt with Opus 5 on a Claude Code project for tightening up my local HomeBridge installation. (HomeBridge is an open-source service that connects a wide variety of third party and otherwise-incompatible smart home devices to the Apple Home platform.) Claude Opus had originally proposed four detailed changes to my code to tidy up my HomeBridge instance, or about three more than I’d bargained for.

After using the “on a budget” prompt, Opus came back with this:

Good constraint — and it changes my earlier answer. I proposed a four-PR program; most of that was thoroughness, not value … Only one of these four issues describes something that is actively costing you anything today 

Yeah, I knew I was being overcharged.

Source link