Project
Expert blog writer
An internal pipeline that turns case notes and video transcripts into evidence-based blog posts for a law firm's five publishing channels. Claude writes; the rules decide what it is allowed to say.
How it works
- Build an evidence pack. For each topic we collect case notes, transcripts and legal sources into one pack.
- Write each channel separately. Claude writes the Naver blog post, the Tistory post and three regional landing pages in five independent calls, through a logged-in Claude Code CLI. Each call sees only the evidence pack and that channel's rules.
- No outside facts. The prompt forbids facts that are not in the pack, and a list of forbidden claims is checked on every draft.
- Review before publishing. A review pass checks that the title answers the reader's question, that the date, court and scope are stated, that key evidence is not missing, and that image captions match the text. A draft with an open legal or media issue is held, even if it scored well.
- Schedule. Approved drafts go onto a publishing calendar. Posts that are already public are protected from automatic edits.
Where it stands
- About 490 completed drafts since July 2026, of which about 300 have been published
- 1,749 automated tests passing
- 173 commits since July 17, 2026
- Every run records its status and estimated model cost, so a failed or rate-limited channel can resume without rewriting the others
We never run Claude Code with permission checks turned off, and the writer does not call outside search. If something cannot be supported by the evidence pack, it does not go into the post.