Feed an agent the code and a recorded conversation; let it write the docs nobody has time to.
The team friction Documentation is the work everyone agrees matters and no one does: tedious, quick to go stale, first cut under pressure. So the team runs on out-of-date or missing docs — and now their agents do too, inheriting the same blind spots. The shared written context the team most needs is the artifact it’s worst at producing.
The pattern The team uses agents to draft and maintain the docs — technical, product, and architecture — from two cheap inputs: the real code, and recorded human context (shoot a video of people talking through the product; drop in the meeting transcript). The expensive part — turning that into a coherent written account — is the agent’s job. The initial set can be generated right out of the Current-architecture workshop. People shift from authoring to curating: review, correct, decide what matters.
Why it works with AI First-pass docs from code plus a transcript sit squarely in an agent’s wheelhouse, and the input is almost free — talking is cheaper than writing. It’s a virtuous loop: the docs agents write become the context that makes the next agent’s work better. The catch is consistency — like a codebase, a doc set agents imitate has to stay coherent, or the contradictions compound (which is why you periodically Compact temporary docs).
Watch-outs Generated ≠ trusted — an unreviewed agent doc is confident and sometimes wrong; the team owns accuracy. Keep the doc types distinct: this is agents producing docs, ADRs hold decision rationale, product memory is curated knowledge agents read from. Generate freely; curate before anyone depends on it.
Related Current-architecture workshop · Compact temporary docs · Consistent, excellent codebase · Product memory