The log ·

A good prompt is just a good ticket

Writing requirements for an AI model and writing tickets for a teammate turn out to be the same skill.

1 min read#ai#workflow#communication

✦ co-written with GLM

I keep noticing that my best AI results come from prompts that look like good issue tickets. Interview me about this observation, then help me condense the pattern into a short log entry.

Drafted from that prompt, then rewritten, cut and checked by hand.

The prompts that give me great results all share a shape with the tickets I write for colleagues: context first, constraint explicit, expected output described, edge cases named. The prompts that fail are the ones I'd be embarrassed to file as a ticket — "make it better", no definition of better.

So I stopped writing "prompts" and started writing tiny tickets:

  • Context: this component renders 500 rows, virtualization is already in place
  • Task: add sticky column headers that respect the existing table skeleton
  • Constraints: no new dependencies, keep the current DOM structure
  • Done when: scrolling keeps headers visible and lighthouse stays green

The side effect surprised me: my actual tickets got better too. Explaining things to a model with zero shared context is practice for explaining things to the new hire with zero shared context.