Prompt Lineage

# Prompt Lineage

This public repo is derived from private working notes and the public blog post:

- https://abhishek.it/blog/never-optimize-the-number

The private notes contained internal traces, exact measurements, customer-adjacent details, and incident-specific examples. Those details are intentionally not published here.

The public prompt keeps the reusable rules:

- measure before believing
- real data over demo prompts
- delivery before latency
- deterministic levers before prompt/model changes
- compile repeated reasoning
- degrade unsupported asks to partial output or refusal
- verify on the real path
- report like an adversary is reading

The repo is meant to be copied into agent optimization work regardless of framework.

If you adapt the prompt for a real system, keep the privacy boundary explicit:

- publish methods, not sensitive traces
- publish failure classes, not customer details
- round or bucket private numbers
- keep exact evidence available internally