research-brief

name: research-brief description: Build a one-page research brief on a single topic for an upcoming podcast episode. Produces fact-cited bullets the host can pull from extemporaneously — never paragraphs of prepared remarks. Each brief includes key facts with working source URLs, three contrarian or non-obvious angles, surprising stats, historical parallels, and named characters with quotes the host can reference. Cites every claim with a verified URL. Run this skill when the user says "research this topic", "build a brief on X", "research brief", "background for the segment on X", "give me ammunition on X", "go deep on T-20260501-01", or any equivalent ask for per-topic prep material. Run once per topic on the rundown. version: 0.1.0

Purpose

This skill is the research desk. The host speaks extemporaneously and does not want to read prose. The brief gives them facts and angles in scannable form — anything paragraph-shaped gets converted to bullets. Three contrarian or non-obvious angles per topic is a hard floor: if the model can only find two, it goes back and works harder. Every factual claim has a working URL the host can click during the recording session if a co-host pushes back.

The brief is per topic, not per episode. A two-topic episode produces two briefs. A three-topic episode produces three.

Inputs

Outputs

Process

When this skill activates, follow these steps in order.

1. Verify config and episode folder exist.

   • If config/podcast.yaml is missing, redirect to podcast-init.
   • If the episode folder does not exist, run episode-program first or accept ad-hoc mode (no run-of-show context).

2. Resolve the topic.

   • If topic looks like a topic ID, load the topic record from the most recent outbox/topics/*.md and the run-of-show. Pull title, hook, source URL, and existing one-line context.
   • If topic is free-form text, treat it as the topic title and proceed without prior context.

3. Set depth parameters.

   Depth	Fact bullets	Contrarian angles	Surprising stats	Historical parallels	Named characters	Source URLs
   light	5–7	2–3	1–2	0–1	1–2	3–5
   standard	10–12	3	2–3	1–2	2–3	5–8
   deep	15+	4–5	4+	2–3	3–5	8–12

4. Gather material.

   • Search the open web for current reporting on the topic.
   • Fetch primary sources where available: company filings, official press releases, founder posts, regulatory documents, transcripts.
   • Pull at least one source on the human side: named executives, founders, analysts, public reactions.
   • Pull at least one historical parallel: a comparable past event, deal, launch, or failure with named participants.

5. Verify every URL.

   • Hit each source URL to confirm it returns 200 and that the page contains the cited claim.
   • If a URL is broken, replace it with a working alternative or drop the claim. Never include a claim with a dead link.

6. Build the brief.

   Required sections in this order:

   • The angle (1 line): the non-obvious lens the host should bring to this topic. Not a summary — a take.
   • Key facts (depth-dependent count): one bullet per fact, each ending with [Source](url). No filler.
   • Three (or more) contrarian or non-obvious angles: each is a one-paragraph bullet — the conventional read, then the counter-read with a fact or two backing it. These are the bullets that earn the host's respect.
   • Surprising stats: numbers the host can drop in. Each cited.
   • Historical parallels: 1–3 past events that rhyme. Named, dated, sourced.
   • Named characters and quotes: the people involved, their public position, one to two quotable lines per person, sourced.
   • What the co-host might ask: 3–5 follow-up questions the co-host could lob in. These are not the same as the run-of-show questions — these go deeper and assume the brief has been read.
   • One thing to avoid saying: optional but useful — a known minefield (legal, factual, or reputational) the host should sidestep.

7. Voice check.

   • Run the generated copy (angles, hooks, framing lines) through style/voice.md plus voice.banned_words_extra from config. Replace banned words. Rewrite passive voice.
   • Factual claims and direct quotes from sources are not rewritten — quotes stay verbatim, attributed.

8. Write the markdown file.

   • Path: outbox/episodes/EP###/briefs/topic-N.md where N is the segment slot.
   • Format follows the section order above. Every fact carries an inline [Source](url) link.

9. Create the Google Doc.

   • If output_targets.google_drive.enabled is true and Drive connector is available, create a doc in briefs_folder_id named EP### — Topic N — <Title>. Mirror markdown content.
   • Optionally share with co-host if the run-of-show is shared with them.
   • Skip if Drive unavailable.

10. Print summary.

    • Topic title, depth, fact count, angle count, source count, broken-link count (always 0 if the skill ran correctly), output paths.
    • Suggested next: run research-brief for the remaining topics on the rundown, then cohost-brief.

Connectors and APIs

Examples

Example 1: Standard-depth brief from a topic ID

User: "research T-20260501-02 for EP042"

Topic: Spotify earnings beat, podcast ad revenue up 28%
Depth: standard
Facts: 11 bullets, all sourced
Angles: 3 contrarian reads
Surprising stats: 3 (ad load %, MAU growth slowdown, exclusive-deal write-off)
Historical parallels: 2 (Pandora 2014, SiriusXM 2019)
Named characters: 3 (Daniel Ek, Dawn Ostroff, Bill Simmons)
Source URLs: 7 verified

Brief written to outbox/episodes/EP042/briefs/topic-2.md.
Google Doc: https://docs.google.com/document/d/xyz789

Next: research-brief T-20260501-01 for the other topic, then cohost-brief EP042.

Example 2: Deep brief on free-form topic

User: "go deep on the WeWork comeback story for the next episode"

Skill accepts the free-form topic, runs at deep depth, builds 15+ facts with 4 angles and 3 historical parallels (1980s real estate flips, Theranos, Yahoo at $5B). Writes to outbox/episodes/EP###/briefs/topic-N.md based on the next available slot in the most recent run-of-show, or asks for the episode number if there is none.

Example 3: Light brief

User: "give me a quick brief on T-20260501-03"

Skill runs at light depth, produces 6 facts, 2 angles, 4 sources. Total length under half a page. Useful for the third or fourth topic in a packed week.