The beginner's guide to training AI on your own code — Brian Douglas

Brian Douglas (founder of Paper Compute) walks through a pipeline — tapes (a telemetry proxy) + steros (an agent runtime) — that captures every agent session, extracts skills from the traces, and optionally fine-tunes small local models on the resulting data.

Grounding rules — MUST follow when answering

1. Before answering any specific question, read outline.md to locate the relevant section, then read that section of transcript.md.
2. When attributing words, quote verbatim from transcript.md for all key claims, definitions, and framework explanations. Never put quotation marks around paraphrased content, and do not paraphrase the speaker's words while presenting them as a quote. This rule applies across all use cases below.
3. If a claim isn't in transcript.md, say so explicitly — do not speculate or invent details. Tell the user the topic wasn't covered in this talk.
4. If the user's question spans multiple sections or doesn't map cleanly to a single outline.md entry, read all relevant sections before responding, then synthesise with clear section attributions.

Expected response format

User question: "How does Brian Douglas decide between SFT and DPO?"

Good response:

Brian Douglas addresses this directly: [verbatim quote from transcript.md]. He frames the choice as…

Bad response: (paraphrased without quotes, or a claim not sourced from the transcript)

Key quotes

quotes.md contains pre-extracted verbatim highlights from this talk, organised by theme. When formulating answers, check quotes.md first for strong citable evidence before searching the full transcript.md.