media-to-roadmap
/media-to-roadmap — external source in, de-duplicated roadmap out
One job: somebody published a video, podcast, article or thread about AI coding, agents or developer workflow, and Ane wants to know what in it is worth adopting. The output is one roadmap file in agent-improvements/, already checked against what the system does today.
Why this exists
The pattern has run by hand at least six times: the Karpathy/Forte video became a 14-item register, the Nate Herk/Cole Medin podcast a 9-item roadmap, the loops video (2026-06-21) the monitoring-loops build plan, the STORM video (2026-06-30) the research-scoping skill, and the Thariq context-engineering article plus the Isenberg graph-engineering video (2026-08-04) a sequenced adoption plan. The pipeline is always the same; the part worth standardising is the judgement call about what NOT to implement, because the source always oversells (STORM's "90 seconds = 48 PhD hours" claim is the reference case).
Method, in order
1. Fetch the source — never work from memory
Every claim in the roadmap traces to fetched content. Retrieval recipes per source type (YouTube transcript with the POT-token fallback, articles, login-gated x.com threads) are in references/fetch-patterns.md. Fetch through context-mode (ctx_fetch_and_index / ctx_execute) so the raw transcript stays out of the conversation. If the source cannot be fetched, stop and say so; do not reconstruct it from training data.
2. Extract concrete techniques
List what the source actually proposes, as specific, implementable claims — not themes. For each claim, record the source's own evidence for it and the strongest objection to it. A claim that survives its objection can be classified; one that does not goes straight to the do-NOT-adopt list with the objection as the reason.