Won't a bigger context window fix this?

A bigger window lets you fit more in — it doesn't make more the right choice. Relevance beats volume; a focused prompt still beats a stuffed one.

Doesn't giving it less make the feature less impressive?

The opposite. Nobody sees your prompt; they only see whether it's right. A narrow, reliable feature earns trust a sprawling, flaky one never does.

Where's the line between "needs a model" and "code can do it"?

If there's one correct answer, code can do it. If a reasonable person would phrase or judge it differently each time, that's the model's part.

Blog / AI agents

Feed Your AI Only What It Needs

The more work you hand a language model, the less you can trust any single thing it gives back.

Kershey Cariño · Jul 5, 2026 · 3 min read

In short

The most reliable AI features give the model the least to do. Hand it only the part that needs real language or judgment, and let plain code own everything exact — numbers, dates, decisions. Feed it the whole world and its quality loosens; give it one clear job and it stays sharp.

The instinct with a new language model is to give it more — more context, more responsibility, more of the job. Reach for the opposite. The AI features I've kept running are the ones where the model does the least.

A model is a probabilistic guesser wearing a very confident voice. Everything you hand it is one more thing it can get subtly, fluently wrong. So the real design question isn't how much it can do — it's how little you can get away with giving it.

give it the job, not the facts

Split the work into two piles: the part that needs language or judgment, and the part that has to be exact. The model gets the first pile. Code gets the second — every time.

Anything with a number, a name, or a date in it is exact. Let a model "summarize the totals" and one day it states a figure that's off, in a sentence that reads perfectly. Let code produce the figure and the model only phrase it, and it can't be wrong about a thing it never calculated.

Hand a model everything, and it does everything a little worse.

less to see, not just less to do

Narrowing the model's job is only half of it. The other half is narrowing what it even looks at.

It's tempting to pour the whole database, the full history, every doc into the prompt just in case. But attention is finite — bury the one line that matters under a thousand that don't, and the answer comes back vaguer, not smarter. Give it the three facts the task needs, not the three hundred it doesn't.

This cuts against the reflex. The flashy move is to let the AI do more and see everything. The move that actually ships is drawing a tight box around the smallest thing it can own.

what needs a model, and what doesn't

For every piece of a feature, ask one blunt question: does this genuinely need a model, or am I handing it over out of laziness?

Anything exact — totals, dates, IDs, a status — stays in code.
Anything with one correct answer stays in code.
The model gets the fuzzy part: phrasing, tone, a call a person would also weigh.

The question was never "how much can the AI do here." It's "what's the smallest piece I can hand it and still get the magic."

Won't a bigger context window fix this?: A bigger window lets you fit more in — it doesn't make more the right choice. Relevance beats volume; a focused prompt still beats a stuffed one.
Doesn't giving it less make the feature less impressive?: The opposite. Nobody sees your prompt; they only see whether it's right. A narrow, reliable feature earns trust a sprawling, flaky one never does.
Where's the line between "needs a model" and "code can do it"?: If there's one correct answer, code can do it. If a reasonable person would phrase or judge it differently each time, that's the model's part.

None of this is trusting AI less. It's spending that trust where it pays — on the one part only a model can do — and refusing to gamble it on the parts plain code already gets right every time.

Seen in a real build

Want it in a real build? A morning report where the model writes only the hello, and code owns every fact. See how MARA · Tickets Reporter is built.

Got something to build?

Got an AI you keep fact-checking every answer? Wire It Right! — contact Kershey, or see mastery.