Prompt engineering basics that still matter
Clear roles, concrete constraints, and reusable templates — practical prompt habits that transfer across GPT, Claude, Gemini, and the rest.
Models change every month. The habits that produce reliable answers don’t. Here’s a short playbook you can apply in any chat — and save into your Vellio library when a template earns its keep.
1. Give the model a job
Start with a role and a deliverable: “You are a staff engineer reviewing a pull request. Return findings ordered by severity.” Vague openers (“help me with this”) waste the first turn.
Browse Coding prompts on Explore for ready-made roles, including Pull request reviewer and Unit test scaffolder.
2. Constrain the output
Say what format you want and what to omit. Bullet lists, tables, “no preamble,” “cite assumptions” — constraints beat adjectives.
For writing workflows, Writing and creative starters like Story premise generator show how a tight output schema keeps answers usable.
3. Put the hard parts in the prompt
Paste the error, the brief, the audience, and the constraints up front. Models can’t invent the facts you withheld. If you’re triage-heavy, try Stack trace triager.
4. Iterate, then save the winner
Treat the first reply as a draft. When a phrasing works twice, save it. That’s the whole point of a prompt library: the second time should be cheaper than the first.
5. Reuse across models
A good prompt is mostly model-agnostic. Run the same template in multi-model chat, keep what lands, and retire what doesn’t. For productivity loops, see Weekly review under Productivity.
Keep going
When you’re ready for model choice rather than phrasing, read How to choose an AI model or skim the live model comparison table.