Technical
- Source materials piped into the draft agent via evidence_refs, not just the synthesized matrix. The current draft agent composes against the matrix, which works for structural claims but underweights operational detail (lifecycle texture, specific code patterns) that lives in the cached source content.
- Internal linking pass. A separate stage reads the final draft, identifies concepts that should cross-link to other KB articles, and generates the markdown link references. Reusable across article types — comparison pieces, deep-dives, tutorials, customer stories all benefit.
Process
- Structured intake stage. Writing piece_brief.md from scratch is the hardest part. The natural product is a two-stage flow: lightweight intake (subjects, URLs, voice anchor), then an auto-generated draft brief the user reviews before the pipeline runs.
- Voice anchor library curation. Editors curate voice samples by structural property; the system picks deterministically against piece metadata. Closes the loop between editorial judgment and automated selection.
Architectural
- Multi-article-type generalization. The architecture supports comparison pieces, deep-dives, customer stories, tutorials, and opinion pieces as configuration changes. Current build proves it for one type. Building two more would prove it generalizes.
- Per-piece source freshness and reproducibility. The corpus is pinned to URLs, not snapshots. A re-run six months from now extracts against whatever those pages say then, which means earlier outputs can't be reproduced exactly. A versioned corpus snapshot would close that.