I have been using Claude Code daily for about a year across multiple repos. These five habits are free, require no plugins, and have saved me more time than any model upgrade.

1. Name your sessions immediately

The default session names are "task-abc123" or "fix-xyz". After three days you have no idea what was in there. The fix is trivial: start every session with a one-line summary.

Bad: claude -> default name task-2026-06-02-abc123
Good: claude -> say "Refactor auth module token validation logic" -> session name is now descriptive

When you run /resume a week later, you will actually know which session to pick.

2. Treat CLAUDE.md as a correction log, not a spec

Most people write CLAUDE.md once and forget it. The better approach is to append a rule every time Claude makes a repeatable mistake.

Examples from my current project:

• "Always use pnpm in this repo. Do not generate package-lock.json."
• "API response keys must be camelCase."
• "Integration tests need a local Redis instance."

Each rule is a single sentence. The file grows organically and future sessions inherit it automatically.

3. Know the difference between /continue and /resume

• /continue (or claude --continue): reopens the most recent session in the current directory. Use this when you stopped working yesterday and want to pick up where you left off.
• /resume (or claude --resume): opens a picker of all historical sessions. Use this when you need a specific older conversation.

The trap is that /resume only helps if you remember the session name or ID. After a few repos and dozens of sessions, the flat list becomes useless.

4. Export decisions before they disappear

When a session contains a design decision, architectural choice, or complex debugging path, run /export before closing it. The output is a markdown file you can drop into your project docs.

The alternative is digging through ~/.claude/projects/ raw JSONL files later. Those files contain everything, but they are not meant for human reading.

5. Separate memory from retrieval

This is the habit that took me longest to learn.

• Memory = reusable facts that future sessions need (CLAUDE.md, rules, conventions).
• Retrieval = finding the old session that actually contained the decision.

CLAUDE.md handles memory well. Retrieval is harder because Claude Code's native tools are designed for exact-session continuation, not for browsing across projects.

After I crossed four active repos, I built a small desktop workspace that groups Claude Code and Codex sessions by project. It reads the same local JSONL files — no cloud, no API keys — and presents them in a browsable project view. I use it to find the right session before I /resume.

https://github.com/Harukaon/shelf

It is not a replacement for /continue, /resume, or CLAUDE.md. It sits beside them as a retrieval layer.

Bottom line

Claude Code is already powerful. The marginal gain from naming sessions, maintaining a correction log, and knowing which command to use when is larger than the gain from most configuration tweaks. And when project count grows, having a way to browse old sessions before resuming them pays for itself quickly.