How to be token efficient using Paul?

[BuggyButLearning]

Whats the best way to be token efficient when using Paul? should i start a whole new session each time issue a new Paul command? or are there certain Paul commands that work better together? Have you thought about using a method like caveman to reduce the token usage?

[cookesan]

I would not start a brand-new session for every Paul command. The framework is designed around persistent .paul state plus managed in-session context, so the efficient path is to group commands around the same phase/plan and only break sessions intentionally.

A good pattern is:

• use /paul:discuss or /paul:discuss-milestone when scope is still fuzzy
• use /paul:plan once you are ready to define the work
• stay in the same session for /paul:apply, /paul:verify, and /paul:unify when they are about the same plan
• use /paul:progress when you want one next action instead of loading more context by guessing
• use /paul:pause or /paul:handoff before a real session break, then /paul:resume in the next session

The docs also point toward keeping plans small: standard plans target a few tasks, and the planning workflow says to load only the context needed for the current plan. That is probably the biggest token saver: do not pull in old summaries or broad project context unless the current plan actually needs them.

So I would start a new session when you are changing milestones, after a completed handoff, or when context is getting too large. I would not restart just between normal commands in the same plan. A compressed note style can help for handoffs, but it should not remove acceptance criteria, decisions, or file references, because those are the parts Paul relies on for continuity.

[BuggyButLearning]

It would be interesting to have Paul enforce the size of plans and even let you know when its time to start a new session. The context window for my projects (mainly game development) fills up quickly and I have to start new session. Also on the Claude pro plan I hit my token limits for the entire day very quickly.

Anything that can be done to have the same results with fewer tokens would be greatly appreciated.