Geoffrey
Shift from managing code files to managing a stateful, visual task board where agents 'raise their hand' for human intervention via a Kanban-based UI.
Key Insights
The 'Raise Hand' UI Pattern
05:11:34Instead of polling agents or scrolling terminal logs, use a visual Kanban board where agents mark cards red to signal they are blocked and require human decision-making.
PR Understanding Quizzes
05:24:49To prevent 'slop' and maintain a mental model, require AI agents to generate a quiz based on their code changes; the human must pass the quiz before merging the PR.
Briefing Books Over Code Diffs
05:28:41Request that agents produce 'briefing books'—high-level conceptual explanations and interactive simulations of changes—rather than raw markdown diffs to increase human-to-AI bandwidth.
The Parity Principle for Agent-Native Apps
05:14:19True agent-native software requires 1:1 parity where every action a user can perform in the GUI is also exposed as a granular tool for the agent to use.
Systems
The Notion-Cloud Code Task Loop
- Record a voice note of feature ideas or bugs.
- Use an AI agent to convert the transcript into structured Kanban tasks in Notion.
- Execute 'notion tasks plan [URL]' in the terminal to have an agent draft a technical approach.
- Review and approve the plan directly within the Notion card.
- Execute 'notion tasks build [URL]' to trigger the code implementation.
- Monitor the Kanban board for status updates or 'blocked' (red) indicators.
Automated Knowledge Transfer Gate
- Agent completes a code task.
- Agent generates a 'Briefing Book' explaining the logic and architectural impact.
- Agent generates a 3-5 question quiz based on the changes.
- Human takes the quiz to verify understanding.
- Human provides a 'vouch' or 'stamp' only after passing the quiz.