Dev.to
7/28/2026
The original title is "Two-stage prompt architecture for AI-generated proposals: a case study"
Original: I Built an AI Proposal Writer in 2 Weeks. Here's the Exact Prompt Architecture That Made It Work.
Short summary
A freelance developer built ProposalAI using a two-stage prompt architecture: first extracting structured data from job postings, then generating tailored proposals from that data. The extraction stage forces the model to understand the client's specific problem before writing anything, eliminating generic output. The stack uses Next.js 14, Supabase, and GPT-4o, with lessons on streaming, caching, and regeneration learned along the way.
- •Two-stage prompting (extraction then generation) eliminates generic AI proposal output
- •Stage 1 extracts structured JSON from job posts; Stage 2 generates proposals referencing the extracted pain points
- •Tech stack: Next.js 14, Supabase, GPT-4o; lessons include streaming from day one and caching extraction results
Generated with AI, which can make mistakes.
Is this a good recommendation for you?
