Note
Why our legal blog writer only writes from an evidence pack
A law firm's blog post can be wrong in ways that hurt a real reader. Here is how we set up Claude so that it writes well and stays inside what we can support.
We write blog posts and landing pages for a Seoul law firm that handles personal rehabilitation and bankruptcy cases. The readers are people in debt looking for an answer. A confident sentence that is slightly wrong, such as the wrong waiting period or a condition that only applies to some courts, can send them in the wrong direction. So the goal is not only to write faster, but to make sure every sentence can be traced back to something we actually have.
One pack per topic
Before any writing starts, we assemble an evidence pack for the topic: our case notes, transcripts of the lawyer's videos, and the legal sources that apply. The pack is the only material the model sees. The writing prompt forbids facts that are not in it, and every draft is checked against a list of claims the firm does not want made.
Five channels, five separate calls
The same topic becomes a Naver blog post, a Tistory post and three regional landing pages. Claude writes each one in its own call with only that channel's rules, so each channel keeps its own format. If one channel fails or hits a usage limit, the run records why and resumes that channel without rewriting the others.
A score is not a pass
A review pass checks that the title answers the reader's question, that the date, court and scope of any rule are stated, that the key evidence is present, and that image captions agree with the text. A draft can score well and still carry an unresolved legal issue, so a draft with an open legal or media problem is held for a person, whatever its score.
What we changed along the way
- We removed fixed style quotas, such as a set ratio of casual sentences, because they pushed the model to fill quotas instead of answering the question.
- We stopped generated posts from being signed as if a lawyer had reviewed them unless that review actually happened.
- We protect every post that is already public, so a rule change never silently rewrites a published article.
Since July 2026 this pipeline has produced about 490 drafts, and about 300 of them have been published. The same approach, a fixed evidence pack and a person at the last step, runs through our case-to-video pipeline as well.