Five agents. One pipeline. A pitch grounded in something your guest actually said.
Paste a LinkedIn URL. HADEF researches every public appearance, extracts full transcripts, finds the angle no one has asked about yet, writes the pitch, and QA-checks it before you see it. The whole thing takes 60–120 seconds.
Agent 1 — Research Discovery
HADEF runs nine searches across podcast directories, conference sites, trade press, YouTube, Spotify, Apple Podcasts, LinkedIn Pulse, Medium, and Substack. Everything from the last 24 months, sorted newest to oldest, deduplicated.
A sufficiency gate checks the result. If there are fewer than two strong sources, the pipeline stops and you are told why. No credits consumed on a prospect who has not said enough in public to pitch from.
Agent 2 — Transcript Engine
For each source, the router classifies the URL type and sends it to the right extractor: YouTube transcript API (instant, free), RSS feed to MP3 to Whisper, Playwright headless browser for JS-rendered podcast players, or BeautifulSoup for article text.
Every successful transcript is cached in Supabase. The second time you or anyone else pitches a guest who appeared on the same show, the transcript is instant. After extraction, a quality classifier marks each transcript as VALID, WEAK, or FAILED. The aggregate coverage score (STRONG, WEAK, or NONE) travels downstream and informs how confident the concept agent should be.
Agent 3 — Concept Development
This is the part that is not a summary. Claude reads the full transcripts as an intellectual portrait and looks for four things: the unexplored thread (a topic the guest returns to but has never been asked to develop), the tension they have named (friction between two competing priorities), the view that has evolved (something they said 18 months ago that sounds different now), and the decision behind the outcome (what the decision logic was, not just what happened).
The output is a concept with a premise, a hook quote with its source URL and date, and the Piar lens: decision quality, cognitive bias, systems thinking, stakeholder logic, or counterfactual reasoning. The hook quote is verified as spoken by the guest, not about them. Confidence is scored HIGH, MEDIUM, or LOW based on transcript coverage and quote origin.
Agent 4 — Pitch Assembly
Two versions: a LinkedIn message (200–280 words, hard limit) and an email variant (280–340 words with subject line). Written in your show's voice. The pitch starts with the specific thing the guest said, not a compliment. No em dashes. No "I am a big fan." No corporate phrasing. The structure is: what they said, the question it opens, why this is the right conversation for your show, and a soft direct ask.
Agent 5 — QA Verification
Five checks before anything reaches your composer:
- Quote accuracy — the hook quote is verified verbatim as a substring of the transcript text, not a semantic match.
- URL liveness — the source URL is checked for HTTP 200.
- Specificity — Claude scores whether the pitch could be sent to a different guest with minor edits. If yes, it fails.
- Coverage cross-reference — if transcript coverage is NONE and a quote is present, the run is flagged HIGH RISK and blocked from auto-passing.
- Confidence check — MANUAL_REQUIRED or LOW confidence concepts are blocked from auto-pass and queued for your review.
QA_PASSED pitches land in your composer with a green check. QA_FAILED pitches land with the failure reason and a recommendation. You decide what to do with either.
What you do
Paste a LinkedIn URL. Click "Find the angle." In 60–120 seconds, read the pitch. If the hook quote feels right, click Send. The pitch goes via LinkedIn DM or email directly from your account. HADEF does not send on your behalf.
That is the whole flow. No CRM to configure. No templates to fill. No summaries to read before writing. Just: paste, review, send.
Questions about the pipeline?
Read the full explainer or talk to us.