Steer AI Writing With Footnotes for Agents

Preserve human intent by attaching context to the exact sentences it explains.

This note covers Bryan Bischof and Adam Conway’s session in the AI Product Engineering series.

Bryan Bischof and Adam Conway are AI engineers at Theory Ventures. The firm’s investors write detailed investment memos, and each memo contains important decisions. When an agent rewrites the prose, these decisions are at risk of getting lost.

Subtext is their fix. It is a lightweight format that attaches metadata to each sentence. The sentence text stays untouched while the author’s clarifications and open questions are stored as metadata. It functions as footnotes for agents so they don’t drift from the source material.

The Subtext skills repo on GitHub.

Humans benefit too. Adam’s demo shows a prototype chat interface where every message surfaces the sender’s linked ticket and call transcript beside the text, so the reader does not have to hunt down the context.

Watch Bryan and Adam’s talk here. Their slides and source repo are also public.

---

For a structured way to evaluate what your AI systems produce, see the AI Evals course, a live cohort with hands-on exercises and office hours.