Prompt engineering steers a model through instructions, examples, and output requirements without changing its weights. It is a fast first step for tasks the model can already handle: development may take hours rather than weeks, but examples add input tokens to every request. section 1.6.2 section 4.0
Build and test a baseline before committing to training:
- Write a clear instruction and provide the context the task requires. Add a few representative input-output examples to demonstrate the desired pattern. section 1.6.2 section 5.1.1
- Specify the output shape. When format is the only problem, a response schema enforced at decode time can produce valid JSON or a fixed template without training. section 1.6.2 section 5.1.8
- Measure accuracy on held-out examples and check run-to-run variability. Fifty examples can reveal a catastrophic gap; distinguishing differences of a few points typically needs 200 to 500. section 1.7 section 4.0
- Inspect errors and revise the prompt, then compare alternatives on the same held-out set. Escalate when the measured quality gap persists, rather than assuming fine-tuning is necessary. section 4.0 section 1.7 section 4.9
Routing support tickets
For twelve-category ticket routing, an eight-shot prompt on Qwen3-4B-Instruct-2507 reached 82% top-1 accuracy on 100 labeled tickets after four rounds of error-driven revisions. The examples were weighted toward high-volume categories; no training pipeline or new checkpoint was required. section 4.0
Prompting cannot supply missing internal knowledge unless that knowledge is included in the request, and it may not reliably enforce behavior the base model resists. Use retrieval-augmented generation for changing documents or verifiable citations; consider fine-tuning when simpler approaches leave a measurable behavioral gap. section 1.6.2 section 5.1.8 section 6.13
Low setup cost does not guarantee low operating cost. In an illustrative eight-shot prompt, about 1,800 input tokens and 50 output tokens cost roughly 0.002X per request, where X is the input price per million tokens and output costs 4X. At high volume, compare that recurring cost with training and hosting; also test whether fewer examples preserve quality. section 4.7.3
Table 4.6 ICL vs. fine-tuning crossover at four volumes (illustrative; units of X per million tokens)
|
Daily volume
|
Daily ICL spend
|
LoRA training one-off
|
LoRA daily hosting
|
Break-even
|
|
100 req/day
|
~0.2 X
|
~200 X
|
~10 X
|
Never (daily hosting alone exceeds API
spend)
|
|
10K req/day
|
~20 X
|
~200 X
|
~10 X
|
About 3 weeks
|
|
100K req/day
|
~200 X
|
~200 X
|
~10 X
|
1 to 2 days
|
|
1M req/day
|
~2,000 X
|
~1,000 X (full SFT)
|
~100 X
|
Under 1 day
|
Compare illustrative prompting and fine-tuning costs across request volumes; recalculate for actual prices and caching. section 4.7.3


















































