coglog tackles one of the most frustrating problems in working with LLM coding assistants: the reasoning that goes into a decision evaporates the moment a conversation ends. You can see the model think through tradeoffs, weigh alternatives, and arrive at a conclusion — and then all of that context is gone, leaving only the resulting code change. This project flips that around by capturing the thinking tokens themselves and structuring them into something durable.

It's packaged as a Claude Code skill that automatically harvests reasoning blocks and files them into three categories:

• Architecture Decision Records (ADRs) — the "why we chose X over Y" rationale that usually lives only in someone's head or a forgotten Slack thread
• Bug post-mortems — the diagnostic path from symptom to root cause, preserved instead of re-derived next time
• Domain logic notes — the implicit business rules and constraints that emerge during implementation

What makes this interesting is the framing: instead of asking developers to write documentation (which never happens), it treats the LLM's own deliberation as the documentation source. The reasoning was already produced — coglog just keeps it. That's a genuinely clever insight about where the friction in project memory actually lives.

This will resonate with anyone who has lost the same architectural debate three times across three conversations, or who maintains a codebase where the rationale behind weird-looking code is buried in old PR descriptions. Solo developers using Claude Code as their primary collaborator will get the most value, but it also has obvious applications for teams wanting to build institutional memory from AI-assisted work without imposing a documentation tax.

The "zero manual effort" promise is bold, but if it delivers even partially, it addresses a real and growing gap in how we work with AI tools.