Call Decoder: a small workflow that shows which questions an earnings call actually answered
Built by PhotonMatrix ·
About this build
A Claude workflow that reads an earnings call transcript and sorts every analyst question into “answered directly”, “partly answered” or “redirected”. It also lists the new facts the call gave compared with the prepared remarks. It’s a reading aid for people who follow company news, not investment advice.
What was built
I got curious after seeing commentators call a CEO’s first earnings call “scripted”. That’s an opinion, and I wanted to see whether a simple tool could back it up or not, using the transcript itself. how it works paste a public earnings call transcript into the first tab the script splits it into prepared remarks and the Q&A Claude takes each analyst question and labels the response: answered directly, partly answered, or redirected to someone else it quotes the key line from the answer so you can check the label yourself a final tab lists anything in the Q&A that wasn’t in the prepared remarks what I learned asking Claude to quote the exact line behind each label made the results much easier to trust and check “redirected” needed a clear definition. I used: the answer is handed to another person, or the question isn’t addressed long transcripts need splitting into chunks, otherwise the labels get inconsistent it’s good at spotting what a call did and didn’t cover, but it can’t tell you what the numbers mean for a stock prompt used (simplified): for each analyst question, label the response as direct, partial or redirected. quote the line that supports the label. then list any new facts in the Q&A that weren’t in the prepared remarks. don’t give investment advice or predictions. known limits it reads only the transcript, so tone and body language are lost labels are a first pass and need a human check it isn’t financial advice and shouldn’t be used to make trading decisions feedback I’d like better definitions for “partly answered” whether a “compare with last quarter’s call” feature would be worth adding how to handle transcripts with poor speaker labels
