How to extract and distribute expert knowledge using AI

Executive overview

Most organisations have critical knowledge trapped in one or two people's heads. When those people leave, that knowledge is gone. The same problem now applies to AI skills: a small number of AI-forward employees are pulling far ahead of their peers.

A three-phase system — extract, package, distribute — pulls expert knowledge out, turns it into AI-ready system prompts, and spreads it across the whole team.

Distributing one person's 3x AI efficiency to 12 teammates creates 36x organisational leverage.

The two symptoms of knowledge concentration

• Tribal knowledge: one expert does a task in 20 minutes at near-100% accuracy; peers take 3 hours at 50–75% accuracy
• The AI gap: AI-forward employees are dramatically more productive, and the gap widens every week
• Both are the same disease: knowledge that benefits one person instead of the whole organisation

Phase 1 — Extract

Three methods for pulling knowledge out of an expert's head:

1. Reverse AI interview — AI acts as a new employee, interviews the expert one question at a time, and produces a detailed SOP at the end
2. Training recordings — transcribe existing training videos and pass the transcript to AI to generate an SOP
3. Documentation audits — consolidate scattered Slack threads, emails, and docs for a topic; AI synthesises them into a draft SOP for the expert to complete

Combining methods improves quality: feed an existing transcript into a reverse interview so the AI asks only about gaps, edge cases, and nuances not covered.

Phase 2 — Package

Turn the raw SOP into a system prompt the AI can operate from:

• Pass the SOP to a new AI chat and ask it to research current prompting best practices for the target model
• Answer three questions: what the AI should do, why (increase leverage by automating this process), and how (define the inputs and expected outputs)
• The AI drafts a deployable system prompt ready for distribution

Phase 3 — Distribute

Three distribution tiers, ordered by user involvement:

1. Prompt libraries — a shared folder (Google Drive, OneDrive) containing the prompt, a use-case description, and a templated version employees can adapt; highest user involvement
2. Projects and gems — tailored AI configurations in ChatGPT/Claude (projects) or Gemini (gems); employees drop in meeting notes or files and get expert-quality output without knowing the underlying prompt; medium involvement
3. Claude skills — autonomous, background-callable tools that trigger from any conversation based on context; near-zero user involvement; other platforms will follow

All three tiers can be used simultaneously. Start with prompt libraries; migrate to projects, gems, and skills as the team grows comfortable.

Recommended first step

This week: identify one bottleneck in your organisation — a skill or piece of knowledge held by one person. Extract it, package it into a system prompt, and distribute it at whichever level suits your team.