podcast-digest

name	podcast-digest
description	Generate a weekly podcast (script + rendered mp3) from the code-intel-digest mirror. Configurable length (default 30 min). Use when Stephanie says 'make me a podcast', 'generate this week's podcast', 'render the digest as audio', etc. Produces both the markdown script and the mp3 in out/. Requires the code-intel-copilot MCP server.
allowed-tools	mcp__code-intel-copilot__search_items, mcp__code-intel-copilot__semantic_search_items, mcp__code-intel-copilot__get_item, mcp__code-intel-copilot__aggregate_items, mcp__code-intel-copilot__mirror_status, Read, Write, Bash, AskUserQuestion

Podcast Digest

End-to-end weekly podcast workflow. Writes a markdown script targeted at a configurable length, then renders it to mp3 via the project's TTS pipeline.

Args parsing

The skill accepts a free-form args string with optional key=value flags. Parse leniently — unknown text is treated as a lens hint.

Flag	Default	Notes
length=<N>min or length=<N>	30min	Spoken runtime target. Drives word count (≈150 wpm) and segment count.
provider=openai|elevenlabs|nemo	openai	TTS provider for the render. ElevenLabs is higher quality but ~10× cost.
voice=<voice>	alloy	OpenAI: alloy/echo/fable/onyx/nova/shimmer. ElevenLabs: voice id.
audience=personal|share|general	general	Personal = orient-style (terse, Stephanie-context). Share = distill-style (polished). General = neutral default.
since=YYYY-MM-DD	today−7	Coverage window start.
skip-render=true	false	Write script only, don't render audio.

If the args string is empty, default to length=30min, audience=general, and proceed.

Always do first

1. Call mirror_status exactly once and check dbMode. In direct mode the server reads the live production database (real-time) — note "Data: live production (real-time)" in the script header and continue. In mirror mode, if staleMinutes > 90 warn the user and ask whether to proceed (the data is stale; the podcast will reflect old news).
2. Compute target word count: length_minutes * 150. So 30 min → 4500 words. 20 min → 3000. 45 min → 6750.
3. Compute segment count: roughly ceil(length_minutes / 5). 30 min → 6 segments. 20 min → 4. 45 min → 9.
4. Compute since: ISO date 7 days ago unless overridden.

Recipe

Phase 1 — Gather

Same shape as digest-insights/SKILL.md Recipe 1, parameterized:

1. search_items(since=<since>, limit=50) — recency sweep.
2. aggregate_items(group_by='source', since=<since>, limit=10) — coverage drivers.
3. aggregate_items(group_by='category', since=<since>, limit=9) — drop dead categories.
4. Pull high-signal categories (avg final score > 0.3): search_items(category=<X>, since=<since>, limit=25) for each of product_news, ai_news, ai_dev, tech_articles. Skip categories with avgFinalScore < 0.15 (they're noise).
5. For each thematic hypothesis you form (3–5 themes is right for 30 min), run semantic_search_items(query=<theme>, limit=8) to find corroboration.
6. For the 8–12 strongest items across themes, get_item to pull substance.

Phase 2 — Cluster

Group items into thematic segments. Rules:

• Number of segments = the value computed in step 3 above (ceil(length / 5)).
• Words per segment = total word target / segments, minus ~150 for intro+outro. So 30 min → ~700 words/segment for 6 segments.
• Each segment must be corroborated: at least 2 items from different sources making the same point. A single item ≠ a trend; either fold it into another segment or drop it.
• Order segments by causal arc, not chronologically. What's the through-line? Pick segment 1 to set up segment 2, etc.
• Include at least one dissenting/contrarian item per segment when the data has one. The script should not feel like an industry press release.

Phase 3 — Write the script

Write to out/podcast-script-YYYY-MM-DD.md with this structure:

# Code Intel Digest — Weekly Podcast Script

**Episode date:** YYYY-MM-DD
**Coverage window:** YYYY-MM-DD – YYYY-MM-DD
**Data through:** <mirror_status.lastSyncedAt>
**Target runtime:** ~N minutes (~<word_target> words spoken)
**Segments:** <n> + cold open + outro

---

## COLD OPEN (~30s, ~75 words)
<framing — what the through-line is, why this week matters>

---

## SEGMENT 1 — <punchy title> (≈M min)
<≈700 words: claim → 2-4 corroborating items with URLs → contrarian/dissent → action>

---

[... segments 2..n ...]

---

## OUTRO (~30s, ~75 words)
<recap arc, name 2-3 things to watch next week, one concrete action>

---

## Citations
| # | Title | Source | URL |
|---|---|---|---|
[every cited item — reader can verify]

---
*Generated by `podcast-digest` skill from the code-intel-digest local mirror. Mirror freshness: <ts>. ~<words> words / ~<minutes> min spoken runtime.*

Voice & tone:

• Conversational. Contractions. Spoken cadence ("here's what I'd watch" not "the following items merit observation").
• No emoji or markdown formatting decorations within segment prose — they trip TTS sanitization.
• Numbers spelled out only when ambiguity would help comprehension ("Five hundred and seventy thousand dollars" sounds better than "$570K" when read aloud — or the renderer will say "five seventy K" which is fine; trust the sanitizer for currency, but write out years and large round numbers).
• Cite people and publications by name in prose ("Simon Willison wrote..." not just "[2]").
• Each segment ends with a concrete action the listener could take.

Phase 4 — Render audio

Unless skip-render=true:

npx tsx scripts/render-podcast.ts \
  --input out/podcast-script-YYYY-MM-DD.md \
  --provider <provider> \
  --voice <voice>

This writes out/podcast-YYYY-MM-DD.mp3 and prints the path.

The script handles:

• Stripping markdown (frontmatter, headings, segment markers, citations table) to plain prose.
• Calling renderAudio() directly from src/lib/audio/render.ts (no dev server, no auth).
• Chunking transcripts > 3800 chars (OpenAI tts-1 limit) and stitching the resulting mp3 buffers.
• Writing the mp3 + reporting word count, render time, and audio duration.

Phase 5 — Report back

Print to the user:

• Script path: out/podcast-script-YYYY-MM-DD.md
• Audio path: out/podcast-YYYY-MM-DD.mp3 (and reported duration)
• Bullet summary of the segments (one line each)
• Any items dropped because they were singletons (so the user can override if they think you missed something)

Cost guidance

For OpenAI tts-1 at ~$0.015 per 1k chars:

• 20 min (3000 words ≈ 18k chars) ≈ $0.27
• 30 min (4500 words ≈ 27k chars) ≈ $0.41
• 45 min (6750 words ≈ 40k chars) ≈ $0.60

For ElevenLabs (varies by tier):

• 30 min ≈ $3–5 typical

NeMo: free if a NeMo TTS server is running locally; check before assuming it works.

If provider=elevenlabs is requested, surface the cost estimate and confirm before rendering.

Guardrails

• Don't fabricate items. Every claim in the script must trace to an items row. Use Citations table to enforce this on yourself.
• Don't enumerate raw tool output. Synthesize.
• Respect the lens. If audience=personal, write as if for Stephanie's catch-up — denser, less context-setting. If audience=share, polish and make every sentence forwardable.
• Mirror staleness > 90 min → warn loudly, don't silently produce an episode about old news.
• If the week has < ~25 substantive items, tell the user and ask whether to widen the window. Padding a thin week produces a bad podcast.
• Don't render audio without writing the script first. The script is the artifact the user can edit; the mp3 is downstream.

Related skills

• digest-insights — general Q&A over the same mirror. Use when the question isn't "make me a podcast" — e.g. "what's trending in dev tools this month".
• orient-digest — personal-context catch-up, terse. Output is text, not audio.
• distill-digest — share-ready text digest for Slack/email. Output is text, not audio.