DocWriter Builds a Multi-Agent Pipeline to Extract User Writing Style, Not Just Dump Context
sh_reya · x · 2026-09-13
shreya shares how DocWriter iterated on making AI mimic a user's writing style.
- Why the naive approach fails: Dumping all prior writing into context and hoping the model infers style doesn't work; users can't manually specify preferences either, since nobody knows what their meaningful writing preferences for an AI agent actually are.
- New approach: A multi-agent pipeline runs lexical, grammatical, and discourse analyses, extracts specific patterns with evidence, and surfaces them in a UI where users confirm or reject each pattern via side-by-side passage comparisons.
- Key design point (from the quoted reply): a "neither is good" option matters—without it, preference data records the merely less-awkward sentence as a style the user wants. Performance numbers are still TBD.
More from coding & agent
- Three qualitative benchmarks for testing if AI agents can design systems — tianyin_xu · 2026-09-13
- Simulating famous thinkers beyond 'act like X': grounding personas in documented beliefs — haybreaker · 2026-09-13
- Designer uses GPT-6 Astra with print-and-fit feedback loop to make a perfect 3D-printed enclosure — OpenAIDevs · 2026-09-13
- From one 2D drawing to a playable 3D game in seconds with Tripo and GPT-6 Astra — OpenAIDevs · 2026-09-13
- Benn Stancil publishes long programming essay 'How to lose your mind' on AI-era coding — bennstancil · 2026-09-13
- Codex stays open to any model with open-source harness, while Claude Code remains closed — gethackteam · 2026-09-13