elumenotion/guideants/session-ingestion-scripting

session-ingestion-scripting

Paths — fixed layout, do not probe or re-derive. The sandbox CWD is the notebook's output directory. This skill's scripts live under Skills/<skill>/scripts/ relative to it, so run the commands in this file exactly as written. Write every deliverable to the CWD with a bare filename (e.g. -o clip, -o scene.wav): never prefix an output path with Output/ — the CWD is the output directory, so Output/… would create a nested Output/ folder.

First step of the video-production pipeline: take a browser-session-capture session (a folder or .zip: video.mp4, narration*.wav, index.json, checkpoints/, event JSONLs, verified edit_map.json) and produce

1. a context report (what was captured, what happened, which audio to use),
2. a timed transcript of the narration (delegated to

audiocpp-timed-transcript), and

1. a TTS-ready script — the story the user told, in their words and order,

cleaned for the TTS engine, with a full audit trail.

The scripting half is a human-in-the-loop editing job. The governing contract is references/tts-editor-brief.md — read it before drafting and treat it as binding. Its first rule: when in doubt, ask; do not draft on a guess. The transcript does not define what the story is; the user does.

Phase 1 — Ingest & report

python3 Skills/session-ingestion-scripting/scripts/session_report.py \
  <session-folder-or-zip> [-o report.md] [--json report.json]

• Stdlib only; read-only on session data. If given a .zip, pass

--extract-dir DIR to unpack first (safe-extract: rejects path traversal).

• The report covers: identity/host/monitor/duration, media (video + narration

candidates with sizes), verified compaction (edit_map.json), idle stats, checkpoint timeline (tab/URL/title/trigger), event kind counts, foreground window intervals, and which checkpoints carry text/screenshot/mhtml.

• Report the bottom line to the user in a few lines: what the session is, what

the narration covers, which WAV to transcribe. Then ask the Phase-2 question.

Phase 2 — Define the story (ask before drafting)

Before any transcript-dependent drafting, get the user's story definition. If the session opener did not supply it, ask exactly this (and nothing else):

In one paragraph: what is the story I told in this recording, and which

parts of the transcript are part of it?

Rules from the brief: 1-3 questions max, numbered, at the end of a short reply; wait for the answer; no "draft assuming X" moves. Log the answer as decided — never re-ask it.

Phase 3 — Transcribe the narration

Route to audiocpp-timed-transcript (timed_transcribe.py) on the narration WAV the user confirms (usually narration.compact.wav — the idle-compacted track that matches the compact video). Keep the outputs: .srt, .vtt, .json, .transcript.txt, plus a derived clean paragraph text. Quote the skill's evidence (engine status, segment/word counts, output files) in the report to the user.

Phase 4 — Draft the TTS script

Draft per references/tts-editor-brief.md:

• Output is ONE working .md file, e.g. <topic>-tts-script.md (in the CWD — no Output/ prefix).

Layout: # TTS-READY … # END TTS-READY spoken block, then Removed (verbatim, one-word reasons), Kept beats, Corrections applied.

• Permitted edits only: fillers, user-identified muttering, flagged

navigation fumbles, user-confirmed ASR fixes, TTS hygiene. Never reorder, summarize, compress, or polish. No length, no structure, no production doc.

• Feedback rounds: edit the same file in place; log every correction.
• Before reporting a script (initial or revised), run:

python3 Skills/session-ingestion-scripting/scripts/verify_script.py \
  <script.md> \
  --removed "<phrase>" [--removed ...] \
  --beat "<phrase>" [--beat ...]

It prints a JSON verdict (open, blockers, warnings, evidence: marker integrity, spoken word count, removed-phrases absent, kept-beats present). Fix blockers; quote the verdict in the reply.

Out of scope

• TTS audio generation (next skill in this set; routes to audiocpp TTS).
• Video assembly, caption burn-in, thumbnails, publishing.
• Any editorial decision about the story — that belongs to the user.

References

• references/tts-editor-brief.md — the binding editor contract (ask-first,

voice preservation, permitted edits, file discipline, anti-patterns).

• references/session-archive.md — anatomy of a browser-session-capture

session folder and which files matter for each phase.

Reporting

End with: the report path, transcript evidence (engine status + counts), the current script file name (unambiguously), and the verify_script.py verdict. If blocked, quote the blocker evidence. No scene tables, no length options, no "suggested next steps" about content the user did not ask for.