Is Claude just not as good at writing documentation/instruction? Also struggling with "reigning it in"
retained text
Something I noticed with Claude models is that they're very good at coding but struggle to document what they're doing and providing instructions for the end-user. It's a bit of an odd-duck for my use-case which is mostly just scripts, wrappers, UIs because things DO work but Claude never produces a complete instruction set like I've seen some of the GPT models, Google models, or even DeepSeek do. It'll just throw a "Run X script with -[settings] and [flags]" without ever explaining to the user what the command does, flags, what's happening behind the scenes, etc. even if instructed to do so. So I end up wasting tokens or effort writing my own breakdown/instructions of the various commands. It's even worse when Claude goes on it's own side-quest outside the scope of the current step and it's all barely documented code and changes for some edgecase that wasn't asked for. My current best efforts at reigning it in has been a "Milestone" approach where it's not allowed to progress to the next part until I've reviewed and approved everything up to a point, but even that doesn't seem to successfully stop it from doing a bunch of undocumented unplanned side-activities that become more and more difficult to track.
document fetched 2026-11-20 00:41 · canonical https://reddit.com/r/ClaudeCode/comments/1y04973/is_claude_just_not_as_good_at_writing