Improved intent accuracy 4× through spec-driven and automated context engineering
Overview
To make AI agents more reliable, I redesigned how we capture and use user intent. By combining spec-driven design with automated context engineering, I achieved a 4× improvement in intent accuracy.
Initial Challenge
- Agents often misinterpreted ambiguous requests.
- Context was inconsistent, making it hard for models to choose the right action.
Approach
- Defined clear, machine-readable specs for what each agent and workflow should do
- Standardized how context is collected, cleaned, and attached to each request
- Automated context building so it is consistent and repeatable
What I Contributed
- Designed the spec format and conventions
- Implemented context engineering patterns that the entire system uses
- Collaborated on evaluation methods to measure improvements in intent accuracy
Results
- Intent accuracy improved by 4× compared to the original baseline
- Users saw fewer misunderstandings and edge-case failures
- Downstream systems became more predictable and easier to monitor