AI coding tools make it easy to generate code, but they also create a quiet documentation problem: useful reasoning gets trapped inside sessions that are hard to find later.
The most useful parts of an AI session are often not the final patch. They are the dead ends, assumptions, trade-offs, and test ideas that led to the patch.
A few practices that help:
- Start sessions from the correct project folder.
- Name work around the task, not the model:
auth-timeout-fix,migration-plan,release-checklist. - Keep important prompts and decisions close to the repo.
- Review old sessions before asking the model to solve a similar bug again.
- Treat session recovery as part of your dev environment, like restoring editor tabs.
I have been building this workflow into an open-source desktop app called Shelf.
Shelf is not another coding model. It is a small Tauri desktop workspace manager for people who use Claude Code and Codex across multiple repos.
The current version can:
- organize sessions by project
- resume previous Claude Code and Codex sessions
- scan local AI conversation history
- restore workspace state after restart
- keep file tree + real PTY terminal in the same window
The main idea is simple: AI coding history should be browsable, not disposable.
Repo: https://github.com/Harukaon/shelf
Related: I also wrote a short workflow note on organizing Claude Code and Codex sessions across projects: https://dev.to/uzoma_uche_3ec83974b4a8a5/organizing-claude-code-and-codex-sessions-across-projects-722
Top comments (0)