Podcast link → transcript / production script package

The big picture

Skill layout                       Data flow
podcast-to-script/
  SKILL.md                         podcast episode link (Spotify / Apple / RSS / --audio)
  scripts/                           │  0 pipeline.py preflight  →  status (find resume point)
    transcribe.py                    ▼
    pipeline.py                    A transcribe.py: fetch + ASR (chunks cached, --workers)
    check_script.py                  ▼
  references/                      $TMPDIR/podcast-to-script/<slug>/   ← episode dir = ALL state
    script-spec.md                   raw.mp3  script.txt  script.srt/vtt  images/  chunks/
    notes-spec.md                    │  B1 agent: outline.md → verify-outline
  tests/                             │  B2 agent: script.md  → verify-script (per-segment resume)
    run_tests.py                     │  C  agent: notes.md   → verify-notes (draft-timeline)
                                     ▼
                                   episode package: raw audio + transcript + outline + script + notes
                                   script.md + notes.md → podcast production pipeline (generate_podcast.py)

Deterministic work is scripted; the agent only does semantic writing (outline /
draft / notes) and runs scripts. Every gate exits 0/1; an interrupted run
resumes at the first stage status does not mark OK.

Resume protocol (do this first)

python3 scripts/pipeline.py preflight            # env & deps OK? fix what FAILs
python3 scripts/pipeline.py status --dir "<episode_dir>"   # or omit --dir to scan all

Continue from the first stage NOT marked OK. Never redo a stage whose gate already passed (transcript chunks are cached; re-runs are cheap and safe).

First run ever (installs deps + warms the ASR model cache):

python3 scripts/pipeline.py preflight --install

Stage A — Fetch transcript (script)

python3 scripts/transcribe.py "<episode_url>"        # ends with: [dir] <episode_dir>

Inputs: Spotify episode URL · Apple Podcasts URL · bare RSS feed (--title picks the episode) · --audio <url> direct file. Options: --model small --lang zh for non-English (default small.en); --workers N parallel chunk transcription (default 4; 1-2 for medium/large models); --proxy; smoke test --max-seconds 180. Performance: small.en ≈ 9x realtime × workers on Apple Silicon.

Gate (script): status shows A fetch OK — episode dir has raw.mp3 (always kept; raw{ext} if not mp3) and non-empty script.txt with char/duration stats. Official <podcast:transcript> is auto-converted to script.txt; episode images (itunes:image + <img> in the description) land in images/ with manifest.json (filename → source URL). Limits: no public RSS = clear error; no speaker diarization (single stream).

Stage B1 — Outline (agent writes, script gates)

Read script.txt, write <episode_dir>/outline.md: episode title line (# {English} -- {中文副标题}) + one segment header per topic (## 第 N 段 · 子标题, same format as script.md) with 3-5 bullet points under each naming the concepts/cases/numbers that segment MUST cover. Follow the original episode's own content flow — the outline is the fidelity audit trail: every part of the source maps to a segment, nothing dropped.

Gate (script): python3 scripts/pipeline.py verify-outline --dir <episode_dir>

Stage B2 — Draft script.md (agent writes, script gates)

Read references/script-spec.md FIRST (full contract: format + quality bar + 旁白 guidelines). Write <episode_dir>/script.md segment by segment, per outline.md. Hard rules (checker-enforced):

Title # {English title} -- {中文副标题}; optional > TTS 制作备注：… quote
Segments ## 第 N 段 · 子标题 matching outline.md
Speaker whitelist ONLY: **主持人**: / **嘉宾**: / **旁白**:
旁白: segment head only, ≤ 4/episode, each ends exactly 好，回到对话。
Fixed closing: ## **主持人**：感谢收听，完整原文列表在 show notes 里。我们下期见
No URLs in any dialogue turn

Quality bar — translate-and-restore first: the script re-presents the ORIGINAL episode's content in Chinese, faithfully covering every argument, example and number — this is a translation of the source show, not new commentary. 主持人/嘉宾 is only the delivery vehicle: host follow-ups exist to draw out content the original already contains; never invent viewpoints or drop arguments. Two-host deep-dive form (open 欢迎收听本期节目, no interview framing); conversational Chinese with English tech terms kept; Chinese punctuation (TTS prosody); follow-ups split into separate turns; transcript has no speaker labels — map the original speakers' exchange onto 主持人/嘉宾 naturally. ≈260 Chinese chars/min for duration.

Gate (script): python3 scripts/pipeline.py verify-script --dir <episode_dir> — also reports which outline segments are still missing, so an interrupted draft resumes per segment. Fix and re-run until exit 0.

Stage C — notes.md (agent writes, script gates)

Read references/notes-spec.md. Write <episode_dir>/notes.md: hook line (本期…。一句话主线：…), 内容速览 bullets, 时间轴, 原文链接 (source URLs live here, never in dialogue). Draft the timeline with:

python3 scripts/pipeline.py draft-timeline --dir <episode_dir>

Images: if images/ is non-empty, reference them in notes.md with Markdown image syntax (after the hook line / near the related bullet, per notes-spec.md). Keep the local filenames from images/ exactly — the publish step matches TOS paths by filename.

Gate (script): python3 scripts/pipeline.py verify-notes --dir <episode_dir>

Deliverable

<episode_dir>/ with raw audio, transcript, images/, outline.md, script.md, notes.md — all gates green. script.md/notes.md are directly consumable by the podcast production pipeline (generate_podcast.py --script … --notes …); keep them together in the episode dir.

Tests

python3 tests/run_tests.py — stdlib-only; run after ANY change to scripts/ or SKILL.md. Covers checker rules, stage-gate transitions, layout end-to-end via file:// fixtures; downstream-parity and real-ASR cases auto-skip when their prerequisites are missing.

Podcast To Script

安装到 Tokeny（自动）

SKILL.md