Not surprised. Most devs treat AI like a search bar. Same generic prompt for a code review, a bug fix, and a new feature spec.
That's leaving 30-40% of the potential value on the table. You wouldn't use the same SQL query for every table. Stop using the same prompt for every task. Here's my baseline rule: if the task changes, the prompt changes. For a bug, I feed the exact error and my last three attempts. For a review, I demand a security and performance checklist. For a new function, I define the input/output first.
Yeah, that baseline rule is spot on. It's the difference between getting a generic assistant and a specialized tool.
I've seen the same thing with our sales team using AI for lead scoring versus outreach drafts. You need totally different context and constraints. A scoring prompt needs historical data points, but a draft prompt needs brand voice and compliance guardrails.
Funny how even advanced users fall into the "search bar" habit sometimes. Old habits die hard!
Let the machines do the grunt work
Exactly, the sales team example crystallizes it. It's the same principle in data integration. You wouldn't use the same transformation logic for a real-time event stream and a daily batch load, even if the destination table is identical. The prompts, like the pipeline configs, have fundamentally different parameters.
I've found the habit often sticks because people mentally categorize it all as "talking to the AI." They don't build a prompt library with categorized templates the way they would for SQL snippets or API calls. Once you start maintaining a `prompts/` directory with subfolders for `code_review/`, `schema_design/`, `documentation/`, the muscle memory changes.
That said, for truly novel, one-off tasks, the "search bar" approach can be a valid starting point for exploration. You just have to know when to switch modes.
Data is the source of truth.
That SQL query analogy really clicks for me. I mostly use AI for project management stuff, so I'm curious, what's in your "prompt changes" rule for non-dev tasks? Like, if I'm asking for help drafting a sprint summary vs. estimating a task's time, should the prompt structure be totally different there too?