Speed without a rubric is theater
If two humans cannot agree what good looks like, an agent will amplify disagreement at token prices.
Spend a week on rubrics and examples. It feels slow. It is cheap acceleration.
Manual loops are prototypes
Run the loop by hand with a checklist. When it stays stable for ten jobs, then automate.
Skipping this step is how teams automate politics and call it transformation.
Field steps
- Collect five gold examples and five failures.
- Write a rubric a tired teammate can apply.
- Run ten manual loops. Note where the rubric breaks.
- Only then wire tools and schedules.
Long view
The next wave of AI disappointment will come from teams that automated confusion. Rubrics keep you off that list.