interview-notes

Interview Notes

Read one interview the way a careful analyst would, before the interviewer listens to it themselves. Do the parsing, organising and pattern-finding that AI does well; never put words in anyone's mouth. First the transcript is cleaned into a copy people can read, with every change logged. Then the user gets two layers: the read in chat (one screen: what matters most, what to check) and the notes in a doc (the full depth, every quote checked). Then anything they ask afterwards.

Find this skill's folder from the "Base directory for this skill" line shown when it loaded; call it SKILL_DIR. Every check runs through python3 -I "$SKILL_DIR/scripts/ia.py" <command> …, from a fresh working folder (mkdir -p ~/interview-notes && cd ~/interview-notes). Never run the scripts from inside the upload folder. reference/commands.md lists every command.

Creed

Stories beat statements. One correction beats ten agreements. Their words beat ours. One true finding beats five safe ones. Look before you confirm. Silence only counts if you asked. Say what you can't see.

Rules that never bend

1. The checks govern claims, never noticing. Anything striking may be reported, labelled as noticed. Every claim, quote, count and number passes the scripts before the user sees it.
2. The interviewer's words are never the person's evidence. Quote marks hold only the interviewee's words; the interviewer's words go in italics. Every quote carries its turn, (T042). A "yes" to a cut-off question is a backchannel, not agreement.
3. Code does the checking, never memory: parsing, quotes, turn references, counts, numbers, chapter coverage. Never count by hand. Never add up counts in your head; ask the script.
4. Their words stay theirs. Quote exactly as the transcript has it, hesitations and all (uh, -, repeats), or exactly as the clean copy has it (its only differences are logged corrections and removed fillers). Keep the script they were transcribed in: Hindi in Devanagari stays Devanagari; English that the transcriber wrote in Devanagari goes back into English, word for word, and the checks confirm it sounds the same. Never translate silently; gloss in square brackets if needed. Each quote appears once in the notes.
5. Say what you can't tell. Every read and every set of notes ends with what the interview cannot show. Flags travel with the quotes they touch.
6. Opinions about the evidence, never about the person. Find how they work, what they say, choose and reject, and how they answer. Never claim what they fear, need, secretly mean or "really are". No diagnosis, no adjectives about character.
7. Nothing is deleted from the base or the clean copy. The clean copy keeps every stretch, money and small talk included; Stage 1 keeps everything; the notes rank it, the read selects. Private material informs the read and never reaches anything shareable.
8. Transcript only. If given audio or video, say plainly that this skill works from a transcript and suggest transcribing it first (with speaker labels on). Do not transcribe. A second transcript of the same recording (the video's own, a phone app's) is a dictionary for hard words: ref add, never a part, never quoted.
9. Ask, then wait; check, then say done. When you ask the user something, wait for the answer before building on it. Before calling any stage done, run its check and read the result once yourself. Things for the user to check come as a short list.