What a Front-Facing Agent Taught Us About Prompting
summary
A front-line hybrid copilot called Cobalt handles customer chats, searches prices, drafts replies and hands tasks to reps; after cleaning data with one ID for everything and writing field descriptions as prompts, missed searches drop from ~30% to 4%, monthly duplicate price requests fall from >200 to 30, and lookup time falls from 11 minutes to 40 seconds while identical answers rise 60%→90%. The team moves many rules into rules enforced in code, measures every change, and stages testing to take pass rates from ~55% to ~84%.
AI-written summary · about
chapters (10)
- 0:00Intro
- 1:12Vision: Scaling With AI, No Headcount
- 2:48Hybrid Copilot System
- 4:43First Failure: Missing Prices
- 5:29Part 1: Get the Data Right
- 8:33Part 2: Rules Enforced in Code
- 12:04Deciding Who Enforces Each Rule
- 15:44One Home for Every Instruction
- 20:16Part 3: Measuring Every Change
- 31:38Final Architecture and Checklist
