Insight tagger
What it does
Tagging qualitative data is coding: attaching a short label to a passage so that everything sharing that label can be pulled together later. Deductive coding applies a codebook you already hold. Inductive coding builds the codebook out of the data as you read. Most projects run both, and over a few months the labels drift unless someone keeps the definitions written down where the team can see them.
This skill codes transcripts, session notes, support tickets and open-ended survey answers. Give it your codebook and it applies that codebook, holding the boundary between near-neighbour codes. Any passage that could sit under two of them gets flagged. Give it nothing and it proposes a codebook first, with a one-line definition per code, and waits for you to approve the set before it tags a single quote. Sentiment is recorded per quote alongside the words that produced the reading.
What you get back
- Tagged passages with the participant and the location in the source, so every tag opens back onto the original.
- The codebook it worked from, with a one-line definition per code.
- Sentiment per quote, next to the phrase that carried it.
- Passages it could not confidently place, held aside for you to look at.
- Counts by participant as well as by mention, so one talkative person can't manufacture a theme.
- A file you can push into Great Question as highlights against the session they came from.
How to use it
Example Prompts:
"Tag these eight transcripts using the codebook in codebook.md."
"Code this batch of open-ends, but show me the tags before you apply them."
"Which quotes did you tag as negative, and what made them negative?"
What it won't do
Sentiment is coarse. Interview politeness reads as positive, and a participant hedging their way through a real complaint often reads as neutral, so use the sentiment column to sort quotes and then read them yourself. It also won't merge overlapping codes without asking you first, which means a messy codebook stays messy until you clean it up.
Questions
Can it use my team's existing codebook?
Yes, and that's the default path. Hand it the codebook in any readable format and it applies those codes only, telling you when a passage fits none of them. Candidate new codes are proposed at the end for you to accept or reject.
Is the sentiment reliable enough to put in a readout?
It reliably sorts a few hundred quotes into piles worth reading. Reporting a sentiment shift between quarters asks more of it than it can give, so use it to find the quotes and let a person make the claim. Every rating keeps its evidence underneath, so you can audit any number before it reaches a deck.
Does it work on survey open-ends as well as transcripts?
Yes, and short open-ends are where it saves the most time. A one-sentence answer arrives with no context around it, so it tags conservatively there and puts anything ambiguous into the unplaced list for you to read.
