There are two very different ways to improve your use of large language models like GPT. One is conceptual and organizational. The other is technical. The first is called a harness. The second is the API.
Let me start with the more important one.
A harness is not a new piece of software. It is simply a disciplined way of using GPT: a set of rules, routines, and constraints you build around the model so it does what you want—rather than “whatever comes to mind.”
Put simply, GPT is like a brilliant, fast person with no agenda, no stable priorities, and no reliable memory. Left on its own, it can produce interesting ideas, but it can also wander, change criteria midstream, exaggerate, or improvise. A harness is what gives it a frame.
Here is a mundane analogy. Imagine someone who cooks extremely well. If you say, “Make something delicious,” you might get a good result—but it will be unpredictable. If instead you say, “We have thirty minutes, these ingredients, four people, and it cannot be spicy,” the outcome improves immediately. Not because the cook got better, but because the problem was defined properly. That is a harness.
With GPT, it is the same. Using a harness means, for example, always specifying the model’s role, what it can and cannot do, the style you want, what must not be invented, and which definitions must remain fixed. It means not firing off prompts at random, but working inside a stable frame. The model is the same; what changes is the discipline around it.
The second way to improve your use of GPT is the API. Here we enter the technical realm, but the core idea is simple. Using GPT via an API means you are not interacting with it as a chat partner. You are embedding it as an engine inside something else: an application, a workflow, or a system.
A simple example: when you use GPT in a browser, you type, wait for the answer, copy, paste, and repeat. With an API, GPT can sit behind a tool that does the same kind of work every time: a system that edits drafts, generates exercises, answers emails, or summarizes documents—consistently and at scale. The API is about automation, repetition, and reliability.
Another household analogy: using chat is like calling someone each time you need help. Using an API is like having that person working in an office—with assigned tasks, procedures, and routines. It is not “smarter.” It is more systematic.
The key difference is this: a harness improves the intellectual quality of what you get from GPT—even for a single person who cannot code. An API improves efficiency and scale, but it does not replace the harness. In fact, an API without a harness often amplifies errors very quickly.
That is why, for most writers, teachers, and professionals, the real leap is not technical. It is conceptual: learning how to build institutions around a general-purpose technology. Almost everything powerful needs a framework—and that framework does not come for free. It must be designed.
To help students build that framework, I will be publishing later this year—together with Raúl A. Sosa—AI at Your Side: The Student’s Guide to Smarter Learning (Oxford University Press).
All these applies not just to GPT, but to large language models more generally.
***If this essay resonated with you, you can subscribe to my Substack for regular reflections on economics, politics, geopolitics, AI, and literature.
*** Disclaimer: I used ChatGPT-5.2. as an editorial and language-refinement tool. The ideas and arguments are entirely my own, and I take full responsibility for them.