Showcase: using scores to find our most expensive and useless prompts

Our LLM spend jumped 30% last month with no feature changes. Traced it to prompt bloat. We used Langfuse's scoring and metadata to find the culprits.

We tag each prompt execution with cost and business context. Then we score for usefulness (e.g., user feedback score, downstream conversion). We query for high-cost, low-score prompts.

```sql
-- Langfuse query example
SELECT
trace_id,
AVG(JSON_VALUE(metadata, '$.estimated_cost')) as avg_cost,
AVG(score_value) as avg_usefulness_score
FROM
scores
JOIN traces ON scores.trace_id = traces.id
WHERE
scores.name = 'user_feedback'
AND traces.name = 'generate_marketing_copy'
GROUP BY
trace_id
HAVING
avg_cost > 0.10
AND avg_usefulness_score < 2.0
ORDER BY
avg_cost DESC;
```

Results:
* A "creative" summarization prompt was costing $0.45 per call. Score: 1.2. It was generating 2000 tokens for internal use where 200 sufficed.
* A customer support Q&A prompt had high variance. Found specific input patterns that triggered long, costly completions with no accuracy gain.

Action taken:
* Fixed token limits.
* Implemented input validation to block wasteful patterns.
* Moved low-priority batch jobs to cheaper models.

Saved ~$2.8k/month. The score-to-cost correlation is the only metric that matters.