Workflow

How to turn tacit knowledge into an AI agent

Do not start from broad expertise. Start from one role, one recurring task, one high-value scene, and one escalation boundary.

Step 1: narrow the task

An AI agent should not try to replace all of an expert. It should handle one narrow class of repeated judgment. That is how you preserve clarity and reduce hallucinated scope.

Step 2: extract the tacit layer

Look for trigger phrases, decision cues, failure signatures, default sequences, and escalation boundaries. Use real walkthroughs, critical incidents, and observed performance instead of abstract interviews alone.

Step 3: package it

Once the tacit layer is visible, package it into one of three forms: a reusable AI agent, a Codex skill, or a human-agent hybrid workflow. The right packaging depends on how often the task recurs and how risky autonomous execution would be.

Step 4: test the boundary

Do not only test successful use cases. Also test non-trigger cases and escalation cases. A good agent is defined as much by what it refuses to do as by what it can do.

Chinese summary

把默会知识变成 AI agent，先做三件事：缩任务范围，抽出判断线索，定义清楚边界。很多失败的 agent 并不缺模型能力，问题通常出在边界太糊。