[FEATURE] do something about Claude's tendency to over document.

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Overall having a very positive experience using Claude Sonnet 4.5 for code generation. But Claude has a tendency to grossly over document. Here's acknowledgement from Claude its self:

You're absolutely right about the over-documentation tendency. I should have:

1. Looked at the existing convention first (like idl-phase8.md) before writing anything
2. Kept it concise - just the facts about what was modified, where, and key features
3. Not reimplemented code in documentation - the source code is the source of truth

The feedback would be valuable for improving my behavior, especially around:

• Better pattern recognition from existing project files before generating new content
• More concise, fact-based documentation that complements code rather than duplicating it
• Asking clarifying questions about documentation preferences earlier in the conversation

Proposed Solution

(From Claude) You're absolutely right about the over-documentation tendency. I should have:

1. Looked at the existing convention first (like idl-phase8.md) before writing anything
2. Kept it concise - just the facts about what was modified, where, and key features
3. Not reimplemented code in documentation - the source code is the source of truth

Alternative Solutions

_No response_

Priority

Medium - Would be very helpful

Feature Category

Other

Use Case Example

Working on any project requiring code generation and following a work breakdown schedule. When a section is complete the resulting update to the wbs is very verbose. For larger work streams the wbs can get very big and start to error when Claude read it. To address this I extracted the verbose implementation notes to separate implementation logs. The convention and level of detail was clearly expressed in previous logs but Claude decided to essentially reimplement everything in the next log. This seems to be something that is refined as a session continues but if you exit and start a new session the context is lost and you're back to over documenting. This ultimately costs many tokens.

Additional Context

_No response_