Stack: Claude (Pro, web or desktop) + the Asana MCP connector One-time setup: 10 to 12 minutes
A commitment made out loud has a short life. Somebody says they will pull the vendor comparison together by Thursday, three people nod, the meeting moves on, and the sentence sits in a notes doc that nobody opens again. Two weeks later you find out it never happened, and the person who said it genuinely does not remember saying it.
The manual fix is that you sit down after the meeting and type the tasks in yourself. Most project managers know this and most do not do it, because it takes twenty minutes and the meeting already ate an hour.
This method compresses that to about four minutes. You paste the notes, Claude applies a fixed extraction rubric you wrote once, and it hands you a review table. You correct the table. Then it creates the tasks through Asana's preview cards, and you approve each one before it lands.
Everybody can already ask a model to "make tasks from these notes." The reason nobody keeps doing it is that the output is bad in a specific way: a 40-minute meeting becomes 23 tasks, eleven of them are discussion points restated as imperatives, four are duplicates of work already in Asana, and six have no owner. You clean it up by hand, decide the cleanup cost more than typing the tasks yourself, and never open it again.
The difference between a tool you use twice and one you use every week is the set of rules that decides what does not become a task. Phase 1 is that set of rules. Everything else is plumbing.
It cannot hear tone. In a transcript, a reluctant "I guess I could take a look at that" is indistinguishable from a firm commitment. It will read both as agreement, and you will have to catch the difference from memory.
It cannot verify who spoke. Automatic transcription tools misattribute speakers routinely, especially in the first minute and on overlapping speech. A misattributed line assigns work to the wrong person with full confidence.
It is not unattended and cannot be scheduled. Asana's write tools render approval cards that require a human click, so this runs when you run it. That is a constraint, and it is also the reason this method is safe to point at a live project.
Enable the Asana connector in Claude. Then, in a new conversation:
List every Asana tool you currently have available. For each one give me the exact tool name and its required parameters. Do not call any of them yet.
Do not skip this and do not trust a tool list you saw in a blog post. Asana's own documentation tells clients to enumerate tools at runtime because the set changes. As of this writing the connector exposes roughly fifteen tools, which is far fewer than some other project connectors, and the names shift between server versions.
Three things to check in that list: