Building in Public: Powering Contract Review with Bad Cases & Dify

Why am I doing this? Most contract-review tools today lean on rigid rule engines that miss edge cases and demand constant upkeep. So over the past few weeks, I prototyped a Dify workflow that marries general checks with a library of “Bad Cases.” Here’s the 4-step pipeline:

• Preprocess & OCR the contract text

• LLM-based initial risk scan

• Retrieve similar Bad Cases from our case bank

• Deep-dive analysis & structured report generation

I’ve just run a demo on 3 sample contracts and 5 Bad Cases—it flags key issues and spits out actionable rewrite suggestions. Along the way, I wrestled with prompt length limits, list-vs-string type bugs, and context-window headaches… Turns out I’m a pretty decent prompt-engineer after all 😂.

Next up:

• Add a loop node for per-case correlation analysis

• Generate clause-level amendment suggestions directly

• Build an interactive feedback loop for deeper dives

What do you think? Any tips on better retrieval strategies or prompt hacks? Let’s co-create—drop your ideas below!