talk-maleix-collective-intelligence

How AI-First Dev Teams Build Collective Intelligence — Edouard Maleix

Edouard Maleix argues that coding agents currently produce lessons every session but the team never learns from them — corrections evaporate into closed chat windows. He proposes a three-act workflow: (1) give agents their own identity and signed commits so work is attributable, capturing rationale in a "diary" of entries; (2) curate entries into thematic "packs" rendered into agent-readable skills that preserve attribution; (3) run controlled evals for fidelity and usefulness, eventually moving toward voluntary task-picking by specialized autonomous agents. The thesis: mistakes should compound into collective intelligence, not disappear into chat history.

Expected bundle files

This skill references two companion files that should be present in the same bundle:

• outline.md — A structured map of the talk, with sections keyed to approximate transcript line ranges and two lookup tables: "Named frameworks / concepts" (e.g. Identity & Diary; Pack, Curation & Render; Evals & Autonomy) and a "Terminology glossary" (e.g. diary, entry, pack, render, fidelity eval, usefulness eval, voluntary task picking, compound engineering).
• transcript.md — The raw speech-to-text transcript of the talk, line-numbered. The bulk is Edouard speaking; a Q&A section at the end interleaves unnamed audience questions with his answers.

If either file is missing, tell the user you cannot ground your answer and ask them to provide the relevant passage directly.

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 short, non-sensitive excerpts from transcript.md. Never put quotation marks around paraphrased content.
3. If a claim isn't in transcript.md, say "the talk doesn't address this" — do not infer positions from outside knowledge.
4. Cite by transcript line range whenever possible.
5. Speaker attribution is unreliable at the sentence level for this transcript — it has no per-speaker labels and includes obvious speech-to-text artifacts (e.g. "Adrian" / "adrian" almost certainly meant "AI"; "globally" likely meant "Claude"; "Tagli" likely "tiny"; "MoltNet" appears once in the abstract but may be transcribed differently in the body). The bulk of the transcript is Edouard speaking; the Q&A at the end interleaves audience questions with his answers. When quoting, preserve the artifacts verbatim — do not silently correct them — but you may note "(likely 'AI')" inline when the meaning is unambiguous from context.
6. Cross-reference any named addressee with the speaker bio / outline.md before attributing. Audience questioners are unnamed; refer to them as "an audience member" or "a questioner."