Drafting the one memory prompt for my Claude + GTD flow: what would you add?
Something I've been wrestling with: I run a GTD-ish setup with Claude wired to my task manager via MCP, and I'm trying to write the one memory prompt that makes the "engage" loop feel natural without being intrusive.
Here's my current draft (swap in whatever tool you use):
>When I complete a task (signaled by "done" / "finished" / "closed it" / "shipped"): – Mark it complete – Fetch the next item on today's list – Surface it as: [project] – [next physical action] plus any blocker note I left – If today is empty, ask whether to pull from tomorrow, drop into someday, or stop
>
>When I'm just talking about a task (no completion verb): – Don't touch the list. Wait for an explicit verb.
>
>When I add a new task via chat: – Confirm title, project, and tags before creating – Default to inbox unless I specify a horizon. Warn me if it's a shared project.
Four things I'm genuinely unsure about and would love takes on:
- Weekly review: anyone baking GTD weekly review into the memory prompt? My instinct is no. The review is the one place I want the friction of doing it manually, and I'm considering explicitly forbidding Claude from running one. But maybe I'm overcorrecting?
- Completion verbs: "wrapped", "killed it", "out the door" all show up in how I actually talk. Am I missing obvious ones? Where does it stop being reliably parseable?
- What to surface next: just title + project, or full context (notes, last touched, related items)? Too much feels like a status report, too little forces follow-ups.
- The "just venting" problem: how do you phrase the "wait, don't act" instruction so the LLM actually respects it? Mine still occasionally jumps the gun on the next task when I'm just complaining.
Curious what people running similar setups actually have written down. Even rough fragments or "this didn't work" examples are useful.
Sharing your prompt verbatim would be much appreciated.