Fighting architecture drift in long ChatGPT projects with an architecture anchor
player_unkn0wn · reddit · 2026-09-26
A Reddit user describes a recurring failure mode in long project chats: the model silently integrates new plausible ideas until the new framing becomes established architecture. Their fix is prepending a self-contained "architecture anchor" to every architecture-related message, listing the fixed design, open decisions, forbidden redesigns, required next-step order, and a drift gate. The core rule: any architecture change must cite new evidence, a real failure, or an explicit user decision—otherwise the model outputs "STOP - ARCHITECTURE DRIFT." An adversarial test (suggesting a "temporary assumption") showed the model correctly refusing to silently operationalize an unresolved decision. They ask the community for stronger approaches: external state files, separate architect/reviewer chats, or periodic re-grounding.
More from coding & agent
- Theo: model routers can't gauge task complexity from prompts alone; escalation-based routing proposed — nbaschez · 2026-09-26
- Code is eating motion graphics: a full launch video in 20 minutes, ~$40, with Opus 5.5 — cedric_chee · 2026-09-26
- OpenClaw 2.0 Hackathon: Build Your Startup's 'First Hire' Agent, Win a Trip to SF and Mac minis — steipete · 2026-09-26
- Claude Opus 5.5 generates pixel-art Nyan Cat animation in under 5 minutes — majidmanzarpour · 2026-09-26
- Users iterate on Opus 5.5 pixel-art animations, one plans to turn it into a game — majidmanzarpour · 2026-09-26
- Agent Flow mod visualizes every Claude Code subagent in a live sidebar tree — EricBuess · 2026-09-26