The Habit That Kept My Prompts Mediocre

For the first few months of using ChatGPT seriously, my prompts looked like this:

"python pandas merge dataframes on multiple columns"

"email subject lines for product launch"

"explain attention mechanism transformers"

These are search queries. Short, keyword-dense, no context, no output format, no target audience. They worked fine on Google. On ChatGPT, they produced technically acceptable responses that rarely matched what I actually needed.

I kept getting outputs that were correct in some generic sense but not useful for my specific situation. The model had no way to know what the data looked like, who the email was for, or how much background the reader had. So it made those decisions on its own, and its defaults were rarely mine.

I didn't notice the pattern until I started writing down what I was actually asking versus what I was getting. The prompts that produced useful responses were structured differently. They had context. They specified who the output was for. They described what I needed the response to do, not just what topic it should cover.

The core difference

A search query is optimized to match documents in an index. Short, precise, keyword-rich. The system does the work of finding relevant content based on those signals.

A prompt is optimized to guide generation. The model doesn't retrieve an answer; it produces one, starting from where your input ends. Every word you write narrows the space of plausible continuations. The more complete your starting point, the more predictable the direction of the output.

This means that almost everything that makes a search query work well makes a prompt work badly.

Dropping context is fine for search, where Google infers what you mean from your history and location. For prompting, the model has no context except what you give it in the conversation. If you're asking about merging DataFrames and the model doesn't know what your tables look like or what you're trying to achieve, it will give you a generic merge example that might not fit your schema at all.

Short and keyword-dense is efficient for search. For prompting, it leaves the model to fill in what the output is for, who will read it, what format it should take, and how detailed it should be. Those are not neutral choices. The model makes them based on what seems statistically plausible, which is often not what you wanted.

Relying on the system to understand intent is the whole design of search. For prompting, intent has to be stated. The model is not inferring what you meant from signals; it is generating from what you wrote.

What changed when I adjusted

The adjustment was less dramatic than I expected. It wasn't a new syntax or a special structure. I started treating prompts more like a brief to a capable colleague who didn't have any background on the project.

The pandas question became:

"I have two DataFrames in pandas. One has customer orders with columns [order_id, customer_id, product_id, quantity]. The other has product data with columns [product_id, name, price]. I need to join them on product_id and return a DataFrame showing each order with the product name and total value (quantity × price). Show me the code with a brief explanation of each step."

That's longer. It's also faster to use in practice, because the first response is useful without two or three follow-up corrections.

The email subject line prompt became:

"Write 5 subject lines for a product launch email. The product is a data analysis template pack for small business owners who use Excel. The tone should be professional but not corporate. No exclamation marks. Each subject line should be under 50 characters."

The attention mechanism question became:

"Explain attention mechanisms in transformers to someone who understands matrix multiplication but has never read a machine learning paper. One paragraph, no equations, one concrete analogy."

In each case, what I added was: who the output is for, what constraints it should meet, and what the specific situation is. None of it was difficult. I had just never thought to include it because I had trained myself on a decade of search query habits.

The tell

The clearest sign you are writing a search query instead of a prompt: you describe the topic but not the output.

"Attention mechanisms in transformers" is a topic. The model can go anywhere with it. "Explain attention mechanisms in transformers to someone who understands matrix multiplication but has never read a machine learning paper" is a direction.

"Email subject lines for product launch" is a topic. "Write 5 subject lines for a product launch email, professional tone, under 50 characters, no exclamation marks" is a direction.

The second version in each pair produces something you can use. Not because it is longer, but because it specifies what success looks like.

The adjustment takes about a week

It felt slow at first. Writing more in the prompt seemed to contradict the efficiency I was used to with search. But the time I spent writing a longer prompt was reliably less than the time I spent correcting a response that had made wrong assumptions.

After a week or two of being deliberate about it, the habit became automatic. I stopped writing keyword-style prompts without thinking about it.

Understanding this distinction, between a query and a prompt, is one of the foundations of what prompt engineering actually is. If you want to build the rest of that foundation systematically, Practical Prompt Engineering covers it from first principles.