Iterating on a Work-Automation Prompt — v1 to v5
The weekly report I write every Friday. It occurred to me that if I built one good prompt, I could save it and reuse it every week — so I spent 4 weeks refining one. I'm recording not just the final version but the failed versions and why they failed. The failures taught me more.
v1 — "Write a weekly report from the notes below" (failed)
I pasted in a week's worth of work notes and let it run. The result looked like a plausible report, but it had two problems: an accomplishment that wasn't in my notes quietly appeared (hallucination), and trivial tasks got the same weight as important ones.
v2 — Added a format spec (half a success)
"Format: ① key accomplishments (max 3) ② in progress (flag risks) ③ next week's plan. Add nothing that isn't in the notes."
The hallucinations disappeared and the structure held. But because the AI was picking the "key accomplishments" on its own, things I considered busywork kept landing in the number-one slot. Judging importance is not something to delegate.
v3 — Marked importance in the input (worked, but tedious)
I starred (★) the important items in my notes and instructed "put ★ items first." The results were good — but starring my notes every week became a chore in itself. The irony of automation creating more manual work.
v4 — Making the AI ask me first (the turning point)
"Before writing the report, pick 5 candidate items from this week's notes that look important, and ask me to confirm them first."
Flipping the order made everything easier. The AI asks, "This week's highlights look like A, B, and C — is that right? Anything missing?" and all I have to answer is one line: "drop B, add D." The manual starring turned into a single exchange.
v5 — Frozen into a reusable template (in use today)
I took v4 and attached the constants (team name, who the report goes to, and one real report from last month as a style example), then saved it. Every Friday: paste template → paste notes → answer the confirmation question in one line → done. Felt time: 40 minutes down to 10.
What 4 weeks taught me
- Make the one-line hallucination guard ("add nothing that isn't there") a default in every automation prompt. The more repetitive the task, the more a single hallucination destroys trust.
- Don't delegate judgment (importance, priority) — delegate the questions that make judgment easy. The leap from v3 to v4 was exactly this shift.
- The refining process is a bigger asset than the finished prompt. I applied the same evolution pattern directly to other recurring tasks (meeting minutes, email).