I've published over 100 articles on Claude Code in the last week. Most were written by Claude Code, humanized, then posted. Here's what the process revealed.
The topics that write themselves
The easiest articles to write: specific patterns with clear before/after. "Rule that causes X problem, fix that solves it." These have natural structure, real examples, and concrete value.
The hardest: opinion pieces with no single right answer. These require a genuine position. Claude defaults to "on one hand... on the other hand" which produces worthless content. Getting real opinions out requires rewriting.
What Claude gets right about technical writing
Structure. Every article came back with clear sections, logical order, appropriate length per section. I almost never had to reorder or restructure.
Completeness. Claude usually covers the angles I'd cover. Occasionally adds one I'd missed.
What Claude gets wrong
Voice. Everything comes back in the same register: measured, helpful, slightly formal, no personality. Fine for documentation, wrong for blog posts.
Opinions. Claude writes balanced assessments. Real developers have opinions. "This pattern is bad and you shouldn't use it" — Claude won't say that unprompted.
Specificity. "This can cause performance issues" instead of "this caused a 400ms delay on a table with 50K rows." Vague qualifications instead of real numbers.
What this means for AI-assisted writing
The draft is easy. The edit is the work.
Claude produces a complete, structured, factually correct first draft in 30 seconds. That draft is also a bit flat, slightly generic, and sounds like Claude wrote it.
The edit — adding real examples, cutting hedging language, injecting an opinion, varying the rhythm — is still human work. The question is whether you do that work or ship the flat draft.
Top comments (0)