Перейти к содержимому

The Missing QA Step in Every Agent Workflow

Intent Solutions

0:00 / 0:00

The Missing QA Step in Every Agent Workflow

154 просмотра · 2 недели назад
Intent Solutions
21 подписчик
154 просмотра · 2 недели назад
Claude drafts four real work emails. Then a second AI audits them blind, and Max has to decide whose read to trust. This is episode two of Intent Solutions' three-part Pain Warp series. Max runs a three-pane (effectively four) Warp setup where Claude, the "Maker," uses a previously built colleague-email skill to draft four emails: a reschedule, an administrator follow-up, a union representation email, and a casual note. He grades every draft blind against a rubric behind a locked decision gate, then runs a separate blind audit in Codex to strip out sycophantic agreement. When the two disagree, Max adjudicates by prioritizing his own voice, the standard he holds every Intent Solutions workflow to. Two of Codex's fixes survive: the reschedule email leaves the next move with the recipient instead of Max, and the administrator follow-up lacks a shared path forward. Max overrules Codex on the union email. He closes by explaining how saving examples and anti-patterns turns the skill into something that keeps improving on its own, the same compounding approach behind every workflow Intent Solutions builds. He previews fuller automation in the next video. This is the kind of AI workflow Intent Solutions builds for clients: real tools, real gates, real judgment calls, not demos. Book a call to talk about one for your team: https://intensesolutions.io Like what you see? Visit maxsheahan.com to check out some of my builds and writing. 00:00 Introduction 01:39 Three-Pane Setup Walkthrough 02:42 The Rubric: Scoring Criteria 05:11 Generating Email Drafts 05:56 Four Email Categories Explained 06:53 Reading the Drafts 09:40 Grading Email 1: Reschedule Collaborator 12:55 Grading Email 2: Follow-up to Administrator 16:08 Grading Email 3: Union Representation 20:26 Grading Email 4: Casual 24:16 Locking Blind Audit & Starting Codex 24:55 Codex Blind Audit Instructions 26:19 Codex Audit Running 28:26 Reviewing Codex Results 30:12 Resolving Disagreements: Marcus Email 34:39 Final Adjudication & Repair Instructions 38:27 Self-Improving Skills & Future Automation 41:08 Outro & Call to Action