AI Podcast Research Assistant — Walk Into Every Interview Knowing More About Your Guest Than They Remember About Themselves
Built by NeonPulse ·
About this build
A Claude and Perplexity powered workflow that takes a podcast guest name and their website or LinkedIn URL and automatically produces a complete pre-interview research package — including a biographical summary, their most interesting and least discussed ideas, 12 conversation-ready questions ranked by depth, topics to
What was built
Built this after listening back to an interview I recorded and realising I had asked questions the guest had answered identically in at least 4 other podcasts I found in the first 5 minutes of research after the fact. The conversation was fine. It was not interesting. The guest had said everything they said to me somewhere else in a better context with a better follow-up question. That was a preparation failure not a talent failure. The workflow runs from two inputs — the guest name and one URL, either their personal website, LinkedIn profile or the URL of something they have written or created. Step 1 — Biographical and career research via Perplexity Perplexity searches for the guest across 6 source types simultaneously. Their own writing — blog posts, essays, articles, books or threads. Interviews they have given in the last 3 years. Talks or presentations they have delivered. Their professional background and career trajectory. Any recent news or announcements involving them. Their social media presence particularly anything they have posted or responded to in the last 30 days. The 30 day social media check is specifically valuable because it surfaces what the guest is thinking about right now — not what they are known for thinking about. People are most animated and most unguarded in conversations when the topic is something they have been actively wrestling with recently rather than something they have a polished answer for. Step 2 — Content analysis via Claude All the research gets passed to Claude with a structured prompt that extracts 6 specific types of content. The core thesis — the central idea or belief that appears consistently across everything they have written or said. This is usually not what they are known for on the surface but the deeper belief underneath their most visible work. The underexplored angle — the topic or idea that appears in their work but that interviewers almost never ask about. This is found by comparing what they write about most versus what they get asked about most in other interviews. The apparent contradiction — any place where something they said or wrote earlier seems to conflict with something more recent. Not a gotcha — a genuine area where their thinking has evolved and where asking about the evolution produces more interesting answers than asking about the current position alone. The thing they are clearly tired of being asked — every prolific guest has 2 to 3 questions they have answered so many times their answer is completely automatic. These get flagged so the host can consciously decide whether to avoid them or to approach them from an angle that breaks the automatic response. Recent developments — anything from the last 90 days that has not been covered in depth in other interviews yet. These are the questions only the most prepared interviewers can ask. Step 3 — Question generation via Claude Claude generates 12 questions structured in 4 tiers. Tier 1 — Opening questions — 3 questions designed to establish rapport and give the guest an easy start while still being more specific than standard openers. These should never be answerable with a simple yes or no. Tier 2 — Depth questions — 4 questions designed to get to the ideas the guest cares about most deeply. These use the core thesis and underexplored angle from the content analysis. Tier 3 — Challenge questions — 3 questions designed to create productive friction without being confrontational. These use the apparent contradiction analysis and approach evolution of thinking as a genuine curiosity rather than a gotcha. Tier 4 — Closing questions — 2 questions designed to produce answers the guest has not given elsewhere. These are the questions that make listeners feel like they heard something they could not have heard on any other podcast. Each question includes a follow-up prompt — a single sentence the host can say if the first answer feels surface level. Step 4 — Briefing document to Notion The entire research package gets formatted into a Notion page with a consistent structure. One page summary at the top that can be read in under 5 minutes. Full research underneath for anyone who wants to go deeper. Questions formatted so the host can work through them naturally during the interview without it feeling like a script. The one page summary contains 4 things — who this person is in 3 sentences, the one idea that makes them worth interviewing, the two questions most likely to produce something nobody has heard before, and the one topic to avoid. What broke during development: The question generation was initially too clever — Claude kept producing questions that were interesting to read but difficult to ask naturally in a conversation. Fixed by adding a spoken test instruction — every question must sound like something a smart curious person would actually say out loud at a dinner table, not something written for an essay. The apparent contradiction detection was generating false positives — flagging normal evolution of thinking as contradiction in a way that could have made the host seem adversarial. Fixed by adding a framing instruction — Claude must frame every apparent contradiction as an evolution rather than an inconsistency and generate the question as a genuine curiosity about how their thinking changed rather than a challenge to their consistency. Perplexity was occasionally surfacing outdated information for guests with a long public profile. Fixed by adding an explicit recency bias instruction — for any topic where old and new information exists always prioritise the most recent. Results after 6 months running this on a weekly podcast: Average pre-interview research time dropped from 3.5 hours to 22 minutes. Three guests have specifically commented during interviews that a question caught them genuinely off guard in a way they appreciated — one described it as the first question in 40 interviews that made her think about something she had not already thought about. One episode produced a clip of a guest reversing a position they had held publicly for 2 years — the clip generated more social media engagement than any other episode in the show's history. That reversal came from a Tier 3 challenge question using the apparent contradiction analysis. The underexplored angle analysis is the single most valuable output. The questions it produces are consistently the ones guests respond to with the most energy because they are being asked about something they care about deeply but rarely get to discuss in public. What I would build next: A post-interview synthesis module that takes the transcript and identifies the 3 most quotable moments, the one idea that deserves a follow-up episode and any threads that were started but not completed — moments where an interesting direction was opened and then the conversation moved on before the idea was fully explored. Will share the full n8n workflow JSON and Claude prompt structure in comments for anyone who wants to deploy this for their own show.