Organize Interview Recordings

Organize Interview Recordings

Invariants

• Keep media local; preserve source recordings, raw transcripts and original notes. Back up only existing files that will change, before writing. New files have no pre-edit backup.
• Reuse the canonical session note and stable filenames. Preserve historical mistakes in the recorded-answer field; correct them only in a visibly purple AI section.
• Existing notes are reuse candidates, not stronger evidence than audio or video. Resolve disputed history from primary evidence; if unavailable, retain the conflict or 未保留完整回答. Local AI-written material never becomes a verified fact merely by being saved.
• AI answers use keyword-led bullets and plain Chinese. Preserve conditions, evidence and responsibility boundaries. No invented historical answers, default experiment gains, metaphors or decorative prose.
• One primary collection owner per semantic question. Full history belongs in session notes; topic collections retain consolidated learning points and source links.
• Search the workspace and user-authorized read-only roots only. An optional untracked local-settings.json beside this file can retain previously authorized roots and runtime command paths; it cannot grant new permission. Never publish it, recordings, notes, resumes, company material, private paths, credentials or authorization artifacts.
• Online sync is a separate, explicitly requested action. No automatic connector setup or online writes during local work.

Choose the smallest workflow

Read each relevant reference once; execute helpers without loading their source unless debugging.

An existing transcript does not require another ASR run. A wording edit does not require rebuilding all sessions or collections. A sync-only task does not require rewriting unchanged local notes.

Local execution

1. Inspect workspace instructions and identify the requested sessions, canonical notes and changed dependencies. Exclude backups, caches and drafts from normal discovery. Use filename/heading inventories first; read relevant sections, not every transcript format or the entire fact library.
2. Reuse valid artifacts. Compare explicit input fingerprints and relevant rule versions with the last successful state. Missing/corrupt outputs, new source evidence or a relevant rule change invalidate the affected stage only. Do not mark a stage complete before validation. Do not globally rewrite legacy notes for a style change unless requested.
3. Back up the selected existing write targets with the helper. Transcribe only missing/invalid source tracks. Preserve overlaps, source labels and raw output; inspect uncertainty windows rather than repeating a full transcription.
4. Read one compact timestamped transcript for reconstruction. Consult raw segments/audio/frames for disputed roles, names, metrics, code and missing turns. Keep all substantive questions and follow-ups; token savings must not silently drop evidence.
5. Write Q&A using the output standard. Give each question a stable explicit anchor and evidence time range when available. End with 考察重点, 主要问题, 改进方向, grounded in this session.
6. Update only relevant primary collection blocks and index entries. If only a source link is new, keep the existing optimized answer. Report inserted/merged/linked/deferred counts separately from source-question counts.
7. Audit changed notes with python scripts/audit_interview_notes.py <notes-dir> --name-regex <selection> --require-purple-ai --require-bullets. Verify the backup manifest and changed collection links/owners separately. The audit checks structure, not factual correctness or semantic equivalence. Resolve failures or explain evidence-dependent exceptions; save successful state last.

Report changed outputs, validation, backup location and unresolved evidence briefly. For requested online sync, include document links, pre/post revisions and read-back results. Do not expose local private details in public release notes.