Front-loading constraints before the request changes AI output a lot
A shift in how to phrase AI requests dramatically cut down on back-and-forth. Previously, a request like "write an email asking my landlord to fix the heating" would come back wrong, prompting a string of corrections: "shorter," "less aggressive," "mention I've already emailed twice" — sometimes six rounds of edits before the output was usable.
The new approach lists every constraint before making the request at all: already asked twice and been ignored, needs to sound firm but not hostile since the tenant still has to live there, must be short enough that the landlord will actually read it, needs a specific deadline, and should imply awareness of tenant rights without explicitly threatening legal action yet. Writing all of that up front, then asking for the email, typically produces a usable result in one pass.
The realization is that all those existed in the requester's head the whole time — they just weren't being stated, and each correction round was really just slowly leaking out information already known. The conclusion: a prompt isn't just the request itself, it's the request plus all the knowledge that makes it specific — and most people's prompts only state the tip of that iceberg.
Key points
- Old approach: short request, then multiple correction rounds (up to six) when the output was wrong
- New approach: list all (tone, length, facts to include, deadline) before the actual request
- Front-loading usually produced a usable result in one round
- Insight: the existed in the requester's head all along, just unstated
- A prompt is really the request plus all the knowledge behind it — most prompts only show a fraction of that