How Your Prompts Change as You Improve — and When Shorter Isn't Better

TL;DR

A prompt isn't just a command — it's a mirror of the mental model you're carrying. Experts write shorter prompts because they've internalized structure. Stage-three overconfidence also produces short prompts — but from skipping constraints. They look identical from the outside. They're different where it counts: one comes with better output, one doesn't. If you're not tracking both, you don't know which one you're in.

Six months into working seriously with agents, I compared two prompts side by side.

The first one: two full paragraphs explaining context, expectations, and examples. The more recent one: two sentences, no examples, no explanation.

I read that as progress. I'd internalized things. The agent "knew" me by now.

Then I looked at the output quality from the last few weeks. Something felt off. Not wrong — just thinner. Less of a specific edge I would have had if I'd done it myself.

The agent hadn't gotten worse. I'd been writing shorter prompts in the wrong way — leaving out constraints I assumed it already knew. It didn't.

01A Prompt Is a Mirror, Not a Command

When you write a prompt, you're not just talking to the agent. You're revealing what you actually think about the task: what matters, what doesn't, what "good" looks like. That tends to be more honest than you'd expect.

Beginners write long prompts because they haven't learned what to include — so they include everything. More experienced operators write shorter prompts because they've learned which constraints are actually load-bearing, and they put those in directly. Less text, but sharper on what matters.

Stage-three operators also write short prompts — but short in the way that leaves out constraints because "the agent knows." These look identical from outside. The difference shows up only in the output.

The beginner Descriptive and long — doesn't know what the agent needs, so provides everything. Output is okay but generic because the prompt isn't specific enough about what actually matters.

The calibrated operator Structural and shorter — knows which constraints are load-bearing and names them explicitly. Compression comes from internalized pattern, not skipped steps.

The stage-three overconfident Also short — but because they've assumed the agent knows things it actually doesn't. Output starts to thin out or miss the specific edge. This person believes they're in the middle column.

Two kinds of "short." Only the output tells them apart.

These three archetypes aren't three different people. They're three phases the same person can move through. And the dangerous one isn't the phase with the longest prompts or the least experience.

02Right Compression and Wrong Compression Both Look Short

Expert compression works because what got dropped wasn't actually needed. Repeated work in a task type teaches you which constraints are the real signal — and you express those concisely without the unnecessary wrapping.

This is right compression. Better output isn't an accident — it's the consequence of understanding what drives quality in this kind of task.

Stage-three compression works differently. Constraints get dropped because you believe the agent already has them from prior sessions or prior context. It doesn't. The agent only has what's in the current session. The things you assume it knows, it has to guess. Output becomes thinner — not wrong, just missing the specific edge you used to get.

The difficulty is that you don't notice. The output is fine. Not clearly wrong — just slightly less of what you're capable of getting. And you attribute that to the task being simpler this time.

03One Habit to Track Your Own Evolution

No framework needed. One small habit, done periodically.

✦

Meta-check: compare a prompt from today with one from three months ago

Pick a recent prompt and a prompt from three months ago for the same task type. Ask three questions: (1) Is today's prompt shorter because I've internalized the pattern, or because I've been leaving things out? (2) Has output quality improved proportionally with the shorter prompts? (3) If I handed today's prompt to someone else, would they have enough context to verify the output? If the answer to two or three is "no" — that's a signal to review what's gone missing.

One habit, three questions, done every few months. Not more than that — just enough to know which archetype you've been operating in lately.

04When Output Quietly Gets Thinner

There's a particular failure mode that's hard to catch: quality declining slowly, with no clear breaking point.

You don't notice because there's no direct comparison. Each output is fine relative to the one before — but compared to work from three months ago, there's been real drift.

The meta-check habit above is built exactly for this. Not for self-criticism, but to have an honest reference point when "feeling like things are working fine" can't tell the difference between actually working fine and being adjusted to working fine.

Prompts are the most honest signal you have about how you're really working. They don't lie — they just require someone who knows how to read them. And the person best positioned to read yours is you, with a bit of deliberate distance.