Juridisk product docs: PRD, roadmap, agent architecture, and AI guardrails

The commit message says "Add desktop project docs," which is telling - these were written offline before any code landed. The repo is being repositioned from upstream willchen96/mike into Juridisk, a Norwegian-language SaaS for AS founders and SMEs priced at NOK 299/mo, with a Stripe-gated free tier capped at roughly 10 queries/day.

The architecture is internally coherent. AGENTS.md specifies a three-agent topology: Router dispatches to either a Legal Assistant Agent (general questions or with document context), a Document Analysis Agent, or a Drafting Agent. All agent calls run parallel Lovdata retrieval first, then hit the rate limiter, then the agent prompt, then stream SSE to the frontend. The caching policy in §2.1 is explicit - search results cached 1 hour, full section text 24 hours, keyed by sha256(query). Chunking strategy in §2.2 is 800 tokens with 100-token overlap; documents under 8,000 tokens get the full text injected; above that threshold it falls back to pgvector similarity search on document_chunks for top-5 chunks.

The CLAUDE.md is a useful artifact on its own: it sets guardrails for AI-assisted development specific to this codebase - things like never logging full document text, always bumping prompt version constants before editing a system prompt in place, and keeping Lovdata full text out of PostgreSQL. That pattern is portable to any fork using Claude for development.

Direct import is unlikely - the PRD is Norway-specific and the pricing assumptions don't translate. But as a reference design for "small-team productization on top of mike," the handoff protocol in AGENTS.md and the chunking/caching policy are worth reading. Dates in the roadmap are aggressive for one person; treat as stated intent.