A call drifts through updates, topics, decisions and side talk in any order. The call report lays it out as a timeline the reader can act on: one section per topic, in the order it was spoken, and one task table on top gathering every task the call produced.
Transcripts are machine-made and wrong in places. Read past the words to what was meant, and say how sure you are.
Process
1. Read the transcript
Ask for the transcript path when the user has not given one. Read all of it.
Look in the transcript's folder for the recording it came from (a video or audio file). With one, every topic's start time becomes a link into it.
Learn who is speaking. Speaker labels in a recorded call are often unreliable: one speaker's audio bleeds into another's track, so a line can appear twice under two labels. Attribute lines from context when the labels disagree with it.
Done when: you have read every line, and you know each speaker's name or role and whether the labels can be trusted.
2. Learn the call's context
A transcript mishears domain words: product names, codenames, people, services. Learn the context the speakers share so you read those words right. The codebase you are invoked from is the call's project:
- Invoke
/project-contextwhen the project has it, and load the references for every domain the call touches. Otherwise read itsAGENTS.mdorCLAUDE.mdand its glossary or domain docs. - Read its recent commits, on every branch, from shortly before the call to now: what was built, what was in flight, and what the speakers were reacting to.
- Do the same for each other repo the call names, when it is on this machine.
Ask the user which projects the call covers when the call strays past the one you are in and you cannot find the others. Context sharpens your reading; the report states what was said, and leaves the state of the code out.
Done when: every proper noun and domain term in the transcript either maps to something you found or is marked unknown.
3. Cut the call into topics
Walk the transcript from start to end and cut it wherever the subject changes. A subject that returns later is a new topic at its new time, with a link back to the first one. Side talk, small talk and stretches the transcription garbled are topics too: keep them, short.
Give each topic:
- from–to: the timestamps it spans.
- title: what it is about, in a few words.
- confidence: a percentage, how sure you are of what was said in this stretch. Low for crosstalk, garbled audio and guessed speakers; high where speakers repeat and agree.
- summary: what was said and meant, attributed to speakers.
- decisions: what was settled, where anything was.
- tasks: the concrete work the topic created, each one a separate checkable item. Write "None" when it created none.
- urgency: for topics with tasks, from what the speakers said about priority.
- notes: anything else worth a label: an outcome, an interruption, where the topic continues.
Refer to people by the names the transcript gives, and as "they" where the transcript does not say otherwise.
Done when: every minute of the call sits in exactly one topic.
4. Gather the task table
Collect the tasks into one table, one row per piece of work, ordered by urgency (most urgent first). A row can merge a topic's checklist into one line. Give each row an owner and the topic it came from. Parked and unclear items get rows too, so nothing said in the call goes missing.
Done when: every topic task is covered by a row, and every row links to a topic.
5. Write the report
Copy template.html to call-report.html beside the transcript.
Fill in the marked constants: the title, the intro, the recording path, tasks and
topics. Replace the sample entries; keep the app code as it is.
The intro is one or two sentences on how far to trust the transcript: label reliability, speakers talking to people off the call, garbled stretches.
Check the script parses, then open the report in the default browser.
Done when: the report opens, and its tables and timeline show every row and topic from steps 3 and 4.
Tell the user the report's path and what lowered your confidence most.