Ten wordings, one problem
Ten people describe the same problem in ten ways, and none of them use your feature's name. One says the export is broken, one asks for a CSV option, one describes waiting for a file that never arrives. Counting keywords finds three requests for a single problem.
The opposite error is worse. Two users say search is slow and mean different things. One means a page load. The other means result ranking. Merging them produces a theme nobody asked for and a task nobody can close.
Grouping by meaning rather than by wording is the whole job. Doing it well means holding the actual texts rather than summaries.
The output is a list of themes, each with its member quotes attached, so whoever prioritizes reads what was actually said instead of trusting a label.
How the ActionFlow runs on the canvas
Tickets arrive through an HTTP Request node from your support tool, reviews through Scrappey, and interview notes through a Document Upload node.
A Text Classifier node assigns each item a coarse type first (bug, request, confusion, or praise), because a confusion report counted as demand distorts everything downstream.
A Loop node iterates over the items within each type. A Generate Object node returns a proposed theme for each item, along with the exact phrase that placed it there.
A Uniq node collapses identical proposals. A second Generate Object pass merges themes whose member quotes describe the same failure, which is where most of the value sits.
Keeping the phrase attached through both passes is the decision that makes the output checkable. A theme with no quotes behind it is a guess.
A Math node counts members per theme. A Sort node orders by frequency. Distinct accounts are counted separately, so ten tickets from one customer do not outrank five tickets from five customers.
A Create Excel node produces the table. A Linear or Jira node opens one item per theme above your threshold. A Slack node posts the summary to the product channel.
The flow prioritizes nothing. Themes arrive as evidence for the person who decides.
Nodes this ActionFlow uses
- HTTP Request (Util) with Scrappey (Util): pulls tickets and public reviews.
- Document Upload (Data): takes interview notes and research write-ups.
- Text Classifier (AI Custom): coarse type first, so confusion is not counted as demand.
- Loop (Control) with Generate Object (AI Core): a theme per item, with the phrase behind it.
- Uniq (Control): collapses identical proposals before the merge pass.
- Math (Util) with Sort (Data): frequency and distinct-account counts, ordered.
- Create Excel (Data): the themes table with member quotes attached.
- Linear (Integration): one item per theme above your threshold. Jira and Monday fit the same slot.
Frequently asked questions
Start building AI workflows
Create a free account, open a template or a blank canvas, and run your first ActionFlow.
Newsletter
Get product updates
New nodes, agents, and product notes. We send mail only when we have something worth opening.