Give AI the Destination, Not the Route

You've probably been disappointed by AI once or twice. You asked for something, and the result was oddly off. Most people conclude, "AI isn't there yet." But the cause is often not AI. It's the habit of the person giving the instructions.

Here's today's principle. You should give AI a destination, not a set of steps. This essay explains why, and how to switch, from start to finish. You can follow along even if you've never used ChatGPT o3. Examples age; principles don't. We'll go slowly.

A good instruction is a coordinate, not a route

Let's start with the principle. Think about how you talk to a taxi driver. "Gangnam Station, Exit 2, please." You give only the destination. You don't recite the route. "Turn left here, go straight at the next light, take the underpass…" Nobody does that. That's the navigation's job. That's the driver's job.

Your job is to pick the destination. Getting there is the driver's job. This is the basis of division of labor.

This common sense has held for centuries, and yet it evaporates the moment AI shows up. People tell AI "first do this, then do this, then do this." When one step goes wrong, they say "AI is dumb." AI isn't dumb. You're the cause — describing a route where you should be naming a destination.

Why this common sense collapses in front of AI

There's a reason. Older AI really did need step-by-step instructions. Two years ago, "draw a picture" gave you nonsense. That's why prompt engineering emerged. You had to stack directions like "4K resolution, cinematic lighting, dark background, 45-degree angle" to get anything usable.

That habit stuck in people's hands. The AI of 2025 is a completely different tool from the AI of three years ago, but the user's way of giving instructions is frozen three years back. Half the reason results disappoint is this. The tool got smart; the user is still treating it like it's dumb.

Example — Escaping the COEX underground with ChatGPT o3

Let's look at a specific example. OpenAI released a reasoning model called o3 in April 2025. Its feature: 'full tool access' — AI chooses which tools to use on its own.

I ran an actual test. The basement of COEX Mall in Seoul is a maze. Anyone who's visited knows. Even the map confuses you. I pasted the COEX underground map image into o3 and asked:

"I want to buy clothes at UNIQLO, then watch a movie at MegaBox. How do I get there?"

Notice what I did not do. I didn't spell out a route. I didn't say "first OCR the map, then extract coordinates, then compute the shortest path." I just gave the destination. UNIQLO → MegaBox.

o3 analyzed the image, web-searched where needed, thought about the route, and returned a 5-minute walking guide like this:

"Exit UNIQLO and face the main corridor. Turn right along the corridor — you'll start seeing KFC signs in a row. Walk straight about 30 meters and you'll reach an intersection with a Spaw tteokbokki shop. Keep straight. You'll pass an Outback Steakhouse…"

I walked it. It was right. The more interesting test came next. I gave a much vaguer goal.

"I have about 2 hours before meeting a friend. I'm not hungry but want a light snack. I don't like sweet things. I want a takeout coffee to drink while walking. Give me a route that fits me."

This is hard to ask even a human. And o3 built a 2-hour course in 15-minute chunks. Iced Americano to-go at the Starbucks by Starfield Library, salty Turkish kebab near the palm trees instead of a sweet snack, a walk through the MegaBox plaza, then Live Plaza. It even told me where the phone charging ports were.

All I did was write one line of destination. AI built the route.

Analogy — Taxi drivers and backseat nags

Think about taxis. A passenger in the back who gives only a destination gets delivered quietly. A passenger in the front seat who barks "turn right here," "change lanes there," "go straight through this light" the whole ride? The driver gets annoyed. Accidents happen. The driver's judgment goes unused.

Your relationship with AI is the same. Give AI only a destination, and its judgment engages. Give AI a narrow route, and it switches off its reasoning and just obeys. One wrong step and the whole result collapses.

One caveat though. If the taxi driver is new to the city, the destination alone isn't enough. Then you add a constraint — "the new cafe near XX Station." Not a step. A constraint. That distinction matters.

Let's confirm with numbers

I ran the same task in three ways. COEX underground: UNIQLO → MegaBox. Compared time and result usefulness.

Instruction style	Prompt length	Result quality
Step-by-step (OCR → locate → route)	~180 chars	Produces a route, but misses shops not on the map
Destination only	~20 chars	5-minute course with shop names on the way
Destination + constraints (time, taste)	~80 chars	2-hour custom course, even phone chargers noted

Instructions cut by 1/9 produced better results. That's the point.

As your words to AI shrink, AI's judgment expands.

How to apply this to your own work

Keep one question in your body.

"Am I giving AI a destination right now, or am I giving it a route?"

The answer splits three ways.

"Only routes" → rewrite

Example: "First open this CSV with pandas, then fill missing values, then group by category, then sum revenue…"

Problem: one wrong step cascades. AI's judgment is unused.

"Destination only" → good

Example: "I want to see monthly revenue trends by category in this CSV."

AI decides whether to use pandas or polars, how to handle missing values. This is enough most of the time.

"Destination + constraints" → best

Example: "I want monthly revenue trends by category in this CSV. One table as the output, ignore seasonality."

Keep the destination intact. Add constraints only for what not to do. Leave AI's judgment room; narrow only the scope you care about.

Here's the mnemonic. Less route. Same destination. Short constraints.

Something you can try right now

Open ChatGPT or Claude. Take a prompt you usually write and shrink it to one sentence.

[drop the long step list]

Destination: [the final result you want, one line]
Constraints: [1-2 things to avoid]

Use this format for a week. At first it feels risky. You think, "is this enough?" It usually is. If it isn't, add constraints one line at a time. The key rule: never add a route. The moment you insert a route, AI's reasoning shuts off.

Summary

Let's wrap up.

Give AI a destination, not steps. Give it steps and it obeys. Give it a destination and it thinks. 2025 AI is a tool capable of thinking. What wakes that capability is how the user instructs.

Solving the COEX underground with ChatGPT o3 is just an example. Next year o5 will arrive; in three years the name will change. But the principle — give the destination, leave the route to AI — gets stronger as AI gets smarter.

Shorter instructions, bigger results. That's the AI playbook for the next 10 years. Technology changes. Principles don't.

Destination only. Route to AI. Constraints short.

[긴 단계 설명 버리기]

목적지: [내가 원하는 최종 결과 한 줄]
제약: [피해야 할 것 1~2개]