AI Prompt Patterns I Frequently Use at Work
I've put together a few prompt patterns I use for planning and documentation work at the office. I've really come to feel that structure matters more than the model.
- Assign a role first: "You are a service planner with 10 years of experience."
- Specify input and output formats: tables, bullets, JSON, etc.
- Include one example: a sample in the tone you want
- Separate requirements from prohibitions: "must include," "never use"
- Run it as draft → critique → revise
The last step was especially effective. Rather than trying to get a great answer in one shot, it was more reliable to generate a first pass and then have the model review it itself.
Prompts are ultimately documents, so when sharing them with the team, it's best to keep a record of versions.
6 answers
Giving the role first really makes a big difference. Just adding one line like "You are ~" changes the quality of the answer.
Do you have a source? If this is just a personal anecdote, it’s a bit weak to use as a template for sharing with the team. In particular, if there were a comparison showing that “draft → critique → revise” is actually better, it would be much more convincing.
Well, it’s true that structure matters, but these days models handle that level of instruction on their own. Aren’t you talking about older models? It feels like overdoing prompt engineering.
Prohibitions really need to be kept separate. If you mix “don’t do X” in with conditions, the model gets confused and ends up doing it anyway lol. I create a separate [Absolutely Prohibited] section at the very bottom.
Good point about version control. Our team makes a separate prompt page in Notion and logs the change history plus a one-line note on "why we changed it this way," which is really helpful later when onboarding new people.
Agreed, especially the last step. I used to try to get it perfect in one shot and ended up disappointed every time, but if I just generate something first and then run "point out just three weaknesses in this answer," it definitely gets better. The catch is that doing this doubles the tokens and time...