Measure before you ship
Start with manual error analysis
Read 20 outputs yourself. You will spot patterns — the model is too verbose, it misses key points, the tone is wrong. This hands-on review builds intuition that automated metrics alone cannot give you.
LLM judges scale evaluation
An LLM judge is another model tasked with grading the output. It can do pairwise comparison (which summary is better?), single-answer grading (rate this from 1 to 5), or rubric-guided grading with detailed criteria for each score level.
Rubrics make judging consistent
A rubric defines what a 5/5 looks like versus a 1/5. For example: 5 = under 100 words, mentions three distinct key points, opens with an overview. You can even give the judge few-shot examples of each score to improve consistency.
Test across models
Platforms like Promptfoo let you run the same prompt against five different LLMs simultaneously and compare results in a table. This helps you choose the right model for your task and spots when one model consistently outperforms others.
If you had to choose between two versions of a summarisation prompt, how would you decide which is better?