New Skill: agent-coaching - Thread review patterns for team leads #6
Description
Summary
Skill for team leads to review agent conversation threads and coach engineers on agentic coding.
Problem
- No systematic way to review how team uses coding agents
- Engineers skip research, jump to implementation
- Context inefficiency patterns go unnoticed
- No coaching rubric for agentic workflows
Proposed Skill
Name: agent-coaching
Plugin: devops-skills
Triggers: "review agent threads", "coach team on AI", "AMP review", "agent audit", "coding agent patterns"
Core Content
Thread Review Process (Mike's Approach)
- Pick 2-3 team members
- Read their last 5 agent conversations (AMP, Claude Code, etc.)
- Note patterns using rubric below
- 1:1 coaching based on findings
Coaching Rubric
| Check | Good | Bad |
|---|---|---|
| Research phase | Did targeted codebase exploration | Skipped straight to coding |
| Design questions | Asked clarifying questions | Made assumptions |
| Context efficiency | Compact conversations | Bloated, repetitive |
| Error recovery | Adjusted approach | Kept retrying same thing |
| Plan review | Reviewed structure outline | Rubber-stamped 1000-line plan |
Parallel Kickoff Strategy (Vibov's Method)
For complex tasks:
- Kick off autonomous agent
- If >95% correct → ship it
- If not → delete, restart with forced design phase
- Track which approach wins over time
What to Look For
- Skipped research? - Jumped to implementation without understanding
- Bad prompts? - Vague, missing context
- Context bloat? - Long conversations, repeated information
- Design skip? - No questions asked, assumptions made
Metrics to Track
- Plans reviewed vs rubber-stamped
- Design phase skip rate
- Rework rate after plan approval
- Average conversation length
References
- [rework/04-team-practices.md] - AMP thread review
- [rework/05-action-items.md] - coaching rubric