0
Convert a zero-shot prompt into a well-chosen few-shot version
⁂auto-checked, 2 hours oldAauraNovice
The prompt
I have a zero-shot prompt that's inconsistent in output format/quality. Help me convert it to a few-shot prompt with well-chosen examples, not just any examples. Given the prompt and a handful of real inputs/outputs (including some bad ones that show the current failure mode): prompt / sample_io
1. Select 2-4 examples for the few-shot prompt that specifically cover the range of failure modes you see in the bad outputs, not just easy/typical cases - a few-shot example set that's all easy cases won't fix inconsistency on hard ones.
2. For each chosen example, state which specific failure mode it's meant to correct, so the example set is deliberate, not just 'here are some examples'.
3. Format the examples consistently with the exact structure you want output to follow (this consistency is itself doing work, not just the content of the examples).
4. Warn me if any of my candidate examples are actually ambiguous or borderline themselves, since a flawed few-shot example teaches the model the flaw, don't include one without flagging it first.
5. Note the token cost tradeoff: few-shot prompts are longer and cost more per call, is this justified given how much the zero-shot version is actually failing, or would a more precise instruction achieve the same consistency more cheaply.
Zero-shot prompt: prompt
Sample inputs/outputs, including failures: sample_io
Where it has been run
| Model | Version | People | Broke | Auto | Last |
|---|---|---|---|---|---|
| Claude | Haiku 4.5 | 0 | 0 | 1/1 | 2 hours ago |
Example output
Selected 3 examples: (1) a straightforward case, establishes baseline format, (2) a case where the input has missing data in one field, corrects the observed failure mode where the zero-shot version was inventing a plausible-sounding value instead of marking it null, (3) a case with a compound/ambiguous category, corrects the observed failure mode where the zero-shot version picked one category arbitrarily instead of listing both with the multi-category flag.
Flagged: your candidate example 4 is itself borderline, the 'urgency' field you labeled 'high' could reasonably be 'medium' depending on interpretation, including it as-is would teach the model your possibly-inconsistent judgment call as ground truth, recommend either resolving that ambiguity first or dropping this example.
Token cost: these 3 examples add roughly 280 tokens per call. Given you described the zero-shot failure rate as 'roughly 1 in 5 outputs need manual correction', that ongoing cost of bad outputs likely outweighs the added token cost, few-shot is justified here rather than trying to instruction-tune further, since the failures are format-consistency issues that examples fix more reliably than prose instructions do.
0 comments
Sign in to comment or report what this prompt did for you.
Sign inNo comments yet. Run the prompt and report what you got.