Turning an AI task list into a Trello board
An assistant that cannot finish a job hands the rest back as prose in a chat, which is a terrible place to keep work. Have it emit the tasks as JSON instead, import them into Trello as cards, and let it read and move those cards later so its picture of what is done stays true without you retyping anything.
Anyone working with an assistant for more than an afternoon ends up in the same place. It did most of the job, and then it stopped, because the last few things need something it does not have: a password only you hold, a decision only you can make, a form that wants a human signature, an approval from somebody in another department.
So it tells you. In prose, in the middle of a conversation that is already two thousand words long. You read it, you mean to do it, and by Tuesday the conversation is closed and three of those tasks are gone.
That is not a model problem. It is a filing problem. The work was captured in the wrong place.
The fix is boring: put them on a board
A chat window is a terrible place to keep a to do list. It has no state. You cannot see at a glance what is left, you cannot reorder it, and nothing reminds you. A kanban board is good at exactly those things, which is why people have used them for decades.
The useful move is to stop reading the assistant's list and start importing it. Ask for the tasks as structured data rather than prose:
Go back through our entire conversation and find every piece of work that is
left for me rather than for you: things you cannot do, things you should not
do, questions you asked that I never answered, and things I said I would do
and never came back to.
Do not include anything you could finish yourself.
Return ONLY a JSON array, no prose around it. One object per task, with
exactly these keys: "Task ID", "Title", "Description", "List", "Labels",
"Checklist".The part that matters is making it go back through the conversation rather than answer from what it happens to remember, and counting a question you never answered as a task. A question left hanging is work you owe just as much as a job you never started, and it is the kind that disappears silently.
What comes back is a file, not a paragraph. Paste it into a Trello importer and you have a board: one card per task, in Backlog, with the steps as a checklist and a label saying why it needs a person.
The half everyone forgets
Getting the tasks out is the easy direction. The hard one is what happens a week later, when you ask the assistant where things stand and it has no idea, because it never saw you move anything.
Most people solve this by retyping. They summarise the board back into the chat, badly, from memory. Then the assistant plans around a picture that is already wrong.
There are two honest ways to fix it.
Let the assistant read the board. Clients that run local MCP servers, which includes Claude Code, Claude Desktop, Codex and Cursor, can be given a tool that reads your board directly. Ask how things stand and it looks, rather than guessing. Tell it a task is finished and it moves the card. The credential for that lives in a config file on your own machine, not in a message.
Or copy the status across. In a plain chat window with no MCP, an export that dumps the board as JSON costs you one paste and needs no credential at all. Less elegant, works everywhere.
Either way the thing that makes it reliable is that each task keeps the id the assistant gave it. Match on titles and the whole scheme breaks the first time somebody renames a card, which is a completely normal thing to do to a card.
Why not just let it run the board
You could give an assistant full write access and let it create and delete cards at will. People do.
We deliberately do not. The tool that reads these boards can move a card and nothing else: no creating, no deleting, no archiving. A board is a record of what humans agreed to do, and a misunderstanding should cost you a card in the wrong column, not a morning reconstructing work that vanished.
Importing is a separate, deliberate action with a preview you check first. That asymmetry is the point.
Trello publishes an official MCP server of its own, free on every plan, and it does cover the whole API including creating and deleting. If you want an assistant that can run your Trello account, install that. The two are not alternatives and plenty of people will want both: a general server works in Trello's own vocabulary of boards and card ids, and the one described here works in the task ids your assistant made up, which is the only name it ever had for the work it handed you.
Doing it
Kanban Importer is a free Trello Power-Up that does the import, tags the cards, and publishes the MCP server that reads them back. There is no account and no card limit. The step by step version has the prompts you can paste and the config for your assistant.
If you are starting from a spreadsheet rather than a chat, the five things that go wrong is the other half of this story.
Common questions
Can I get ChatGPT or Claude to make Trello cards for me?
Yes. Ask for the task list as a JSON array rather than prose, then paste it into a Trello importer that accepts JSON. Each object becomes a card, with its description, checklist and labels already attached.
How does an AI assistant know when I have finished a task?
Only if it can see the board. Either give it a tool that reads Trello directly, which clients running local MCP servers can do, or export the board status and paste it into the chat. Without one of those it is working from whatever you last told it.
Should an AI assistant be allowed to delete Trello cards?
We think not. Reading and moving a card is useful and recoverable. Creating and deleting is where a misunderstanding costs you real work, so the tool published with Kanban Importer cannot do either.
What stops the assistant moving the wrong card?
Each task keeps the id the assistant invented for it, written both into the card and into the Power-Up's own storage. It matches on that rather than on the card title, so renaming a card or adding a note to it changes nothing.
Is this the same as the official Trello MCP server?
No, and they work together. Trello's own server covers the full API and is the better choice for general work in your boards. This one does a single thing it cannot: it addresses cards by the task id your assistant invented, so "mark the Stripe webhook task done" works without anybody looking up a card id, and keeps working after the card is renamed.
Does this cost anything?
No. The Power-Up, the import, the export and the MCP server are all free, and there is no account to create.
Need something like this built around your own workflow? That is custom app development.