Relationship Debrief

Rami works with a small team and wants to get better at communicating with each specific person. The point of this skill is to make every call leave a trace: a clear read on each person, and a concrete way to show up better with them next time. Done repeatedly, these traces accumulate into a real understanding of each teammate in his My Social Circle/ CRM.

This is the relational counterpart to communication-analysis. That skill grades how Rami spoke (vocab, clarity, persuasion). This skill is about the people on the other side and the relationship between them and Rami. Run both when he wants the full picture; run this one alone when he cares about the people, not his delivery.

What you produce

1. A debrief report - the artifact Rami reads. Always written, always safe. Goes to <transcript_dir>/debriefs/<transcript-name>.md.
2. Proposed CRM updates - ready-to-apply blocks inside the report (durable observations per person, a meeting note, follow-up tasks). These touch the vault only when Rami confirms.

Default to writing only the report. Apply changes into My Social Circle/ and My Meetings/ solely when Rami explicitly asks. The vault folder is gitignored with no undo and holds real data about real coworkers - that asymmetry (a wrong auto-edit is unrecoverable, a proposal costs nothing) is why proposals are the default.

The hard part: who actually said what

Read references/sources-and-schema.md before doing anything else. The input format (OpenOats) has a trap: it records two audio streams, and the You channel (Rami's mic) echoes other people's words a few seconds later. So You is not reliably Rami, and the anonymous Speaker N labels mean nothing across calls.

Getting attribution wrong poisons everything downstream - you would record observations about David that were actually Dan. So treat identity as an inference problem, not a given: