The original title is "How we built a smart AI writing agent for blog articles"

Short summary

The author built 'Larry,' an AI writing agent for lawyers that avoids the common pitfall of full-document regeneration by using surgical edit tools (replace_in_field) that patch specific text while preserving everything else. The agent follows a strict interview-first loop: it researches the topic, asks the expert for a concrete anonymized case, proposes a structure, then writes and self-critiques. Key patterns include a validation gate the model can't fake, external URLs sourced only from web search, and no agent framework—just a while loop with three tools.

• •Surgical replace_in_field tool prevents voice/fact drift by editing only targeted text instead of regenerating whole articles
• •Interview-first loop captures domain expert's real cases and anecdotes before any writing happens
• •Validation gate forces model to re-read article before certifying an example exists; external links come from search only

Generated with AI, which can make mistakes.