The root llms.txt compresses most of its instructions and product links into one extremely long line, which is poor for line-oriented chunkers, diffs, citation systems, and humans debugging AI behavior. It also lists starter pages without saying when to pick each product.
Two phases. The mechanical fix — every entry on its own line — can ship immediately. The restructure around tasks (build an AI voice agent → Agents namespace, control a call in real time → RELAY, declarative call control → SWML, including what not to start with, e.g. the Compatibility API for AI features) should wait for #562 and derive from its manifest rather than encoding routing rules here independently. Audit §4.
Note: the root llms.txt is generated by Fern, and past attempts to configure its output via docs.yml have hit limits — part of this may be an upstream Fern request.
The root llms.txt compresses most of its instructions and product links into one extremely long line, which is poor for line-oriented chunkers, diffs, citation systems, and humans debugging AI behavior. It also lists starter pages without saying when to pick each product.
Two phases. The mechanical fix — every entry on its own line — can ship immediately. The restructure around tasks (build an AI voice agent → Agents namespace, control a call in real time → RELAY, declarative call control → SWML, including what not to start with, e.g. the Compatibility API for AI features) should wait for #562 and derive from its manifest rather than encoding routing rules here independently. Audit §4.
Note: the root llms.txt is generated by Fern, and past attempts to configure its output via docs.yml have hit limits — part of this may be an upstream Fern request.