Claude Code Task Management: A List Is Not a Plan

You ask Claude Code to add team invites. It writes itself a task list: schema, endpoint, email template, settings screen. Three ticks later you close the laptop.

Next morning the list is still there. That part works now. What is not there is the reason you dropped the settings screen on Tuesday, or the fact that the endpoint it ticked off was built against a decision you reversed on Wednesday. The list kept its place. The project moved underneath it.

What the built-in tasks do well

Older guides will tell you Claude Code's to-do list dies with the session. That was true of the original Todo list. Tasks replaced it, and per the third-party guides we read (ClaudeLog and ClaudeFast, not Anthropic's own docs, so check the changelog for your version) they are saved as files in your home directory, survive a restart, and can be shared between sessions by giving each one the same task list ID. One task can also block another.

For a feature that takes an afternoon or a weekend, that is the right tool. Use it. The rest of this post is about what happens when the work stops being a feature and becomes a project.

Where a list runs out

It answers what, not why. A task says “build the invite endpoint”. It does not say that invites are email-only because you decided against shareable links. A new session reads the list faithfully and can still undo the call.

Done means the AI said so. A task is completed when Claude marks it completed. One of the guides is plain that checking the work first is an instruction you add to CLAUDE.md, not something the list does on its own. Nobody compares the result with what you meant.

It is flat. Tasks and dependencies describe the next hour. Nothing on the list asks whether the next three weeks still point at the right thing, or notices that the plan has grown since you wrote it.

It lives with one tool on one machine. Move part of the project to Cursor, or open it on another laptop, and the list does not come along.

A task list holds work. A plan holds intent.

The difference shows up around week three. The list is accurate about what is pending and silent about what has changed. You end up as the only place the project's direction is written down, and you are the one who forgets.

A list remembers where you were. A plan knows where you are going.

What a project needs on top of the list is small. A written reason next to each decision, marked replaced rather than deleted when you change your mind. A check for done, written before the work starts. A short note after each build saying what actually happened, including the things that were never on the list. And a human saying yes, no or not yet.

You can do all of that in a plain document, if you keep it up. Most people stop keeping it up the first busy week, which is the week it matters.

One way to run it

That layer is what we built PAPI for. PAPI is a project manager that works inside your AI chat, so Claude Code reads and writes it as part of the build. You keep the tasks you already use for the small stuff.

Work runs in cycles, usually 3 to 8 tasks. Each task gets a build handoff: the goal, what is out of scope, how to check it is done, and the risks, written so an AI with no history on the project can run it. Every build files a report, you accept, request changes or reject, and decisions keep their reasoning. The setup for Claude Code is on the Claude Code page.

When to stay with the built-in list

If the work fits in a week and you will be the only one touching it, the built-in tasks are enough, and anything more is paperwork. If you throw away most of what you build, a plan is overhead for now.

PAPI does not write code, and it does not make the list obsolete. Your assistant still builds, and you still decide what is good enough.

A test for your own project

Open a fresh session and ask what you decided last week, and why. If the answer comes from your head and not from the project, that is the gap. We covered the session side of it in Claude Code Loses Context, and the board side in Project Management for AI Coding. Walk one cycle on a demo project, no signup needed.