Proposal: Supporting a Human-Authored "Anchor" Problem Statement in Workflows

[yogiadi]

Hey everyone,
I've been thinking a lot about a common problem in AI-assisted development workflows (including tools like OpenSpec). Right now, standard flows often allow users to prompt an LLM from a blank slate to generate complete specification documents. However, even the best engineers cannot solve a problem if they lack context. Users are often writing a single one or 2 liner prompt which gets converted to full spec open for review by author. Since it doesn't have all the context from those 2 liner prompt it fills in missing architectural and business gaps with plausible-sounding hallucinations. Because of the volume , it is not reviewed thoroughly by the users.
I believe we need to anchor AI generation in reality by encouraging or supporting a Human-First Seed Pattern. A human engineer writes a foundational problem statement/context document directly (e.g., minimum 50 lines detailing the actual pain point, boundaries, and trade-offs) before triggering downstream spec generation. This document cannot be modified by AI directly and will have a write lock on this. AI acts as an expander, ingesting the human anchor as its primary source of truth to flesh out tasks, design documents, and API contracts, rather than inventing the core problem statement itself.
How are other teams handling context drift when generating specs with AI? Do you think it makes sense for tools like OpenSpec to support native validation or configuration rules that check for a human-authored anchor file before proceeding?
Would love to hear your thoughts on whether this aligns with the project's direction!