How I AI: How Tomasz Tunguz digests 36 weekly podcasts without spending 36 hours listening

Tomasz Tunguz follows 36 podcasts, but he doesn't have 36 hours a week to listen to them. Instead, he built a local system that downloads new episodes, transcribes them, extracts topics and company mentions, and prepares research and writing drafts for his review.

As the founder of Theory Ventures, Tomasz writes frequently about enterprise software and investing, and this system feeds his research.

He prefers reading to listening because he can scan and skip ahead, so he built a searchable pipeline to convert the audio he follows into text he can use.

The terminal-based tool, which he calls the Parakeet Podcast Processor, is designed to fit his personal workflow like a glove, rather than serving as a general podcast product.

In this episode of How I AI, Tomasz explains the transcription, storage, extraction, and writing stages of his hyper-personalized software, including the prompts he continues to revise.

The Parakeet Podcast Processor: From audio to searchable notes

Every day, the tool checks 36 podcast feeds, downloads new episodes, and turns the audio into searchable transcripts and structured notes.

The process runs locally on his Mac:

• Download and transcribe: `ffmpeg` prepares the audio, then a speech-to-text model creates the transcript. Tomasz began with OpenAI’s Whisper and, at the time of recording, was using Nvidia’s Parakeet locally.
• Clean transcripts: Gemma 3, running locally through Ollama in the demonstrated setup, cleans the raw transcript with a prompt that asks it to remove ums and ahs while preserving all technical conversations.

You're a transcript editor. Clean up this podcast while preserving all the content. Keep the same length, remove the ums and the ahs, preserve all technical conversations.

• Orchestration and storage: A podcast orchestrator script runs the processing steps, and a local DuckDB database stores transcripts and tracks processing status.
• Daily extraction: The system sends the day's cleaned transcripts through a prompt that produces a structured summary for Tomasz to review.

For each episode, the structured output includes:

• Host and Guest: Quick context.
• Comprehensive Summary: A high-level overview.
• Key Topics and Themes: Categorized insights from the conversation.
• Actionable Quotes: Specific snippets that resonate or spark ideas.
• Investment themes: The model suggests companies or markets that may deserve further research, which can kick off a market-mapping project. These are leads, not investment recommendations, and Tomasz returns to the source material and his normal diligence before making decisions.
• Possible observations: The system drafts ideas that may become social posts. Tomasz notes that he is still refining the prompt for this section.
• Company mentions: The system identifies companies mentioned in a transcript so Tomasz can research them further. He initially tried the Stanford named-entity recognizer but found it struggled. After trying to improve performance by cleaning up the transcripts, he found that simply sending the text to a larger language model worked much better.

Why a terminal is the right UI

The screenshots from this episode show a terminal interface, not a separate graphical application.

Tomasz values the terminal’s low interaction latency, citing Dan Luu’s writing on the topic as an inspiration. He began using terminal tools more heavily during the COVID period and now uses them for everything from email to custom scripts.

A terminal interface is a good fit for Tomasz because the commands are fast and easy to connect to scripts. For another user, a graphical interface might be better; the useful design is the one that reduces friction for the person operating it.

Because the tool is personal, Tomasz can change a script whenever his feeds, schema, or review process changes. The podcast-processing workflow shows the download, transcription, storage, extraction, and review loop behind that glove-like fit.

Using AI as an AP English teacher

Tomasz doesn’t just use AI to consume information. He also uses it to turn ideas from his podcast summaries into rough blog drafts.

From podcast insight to blog draft

1. Contextual Generation: Tomasz uses a "podcast generator" script. He feeds it the transcript of a relevant podcast (or just an idea) and a prompt outlining the desired content.

You are an expert blog writer specializing in technology and business content.
Based on the style analysis of existing posts, write in a style that:

f''.join(style_analysis.tone_characteristics)}
Uses paragraphs of approximately {style_analysis.avg_paragraph_length} words
Follows these patterns: {f''.join(style_analysis.common_patterns)}
Hook examples from existing posts:

{chr(10).join('+' + hook['100'] + ' ' for hook in style_analysis.hook_examples)}

CRITICAL REQUIREMENTS:

Write approximately 500 words total
NO section headers or H2/H3 tags - write as continuous flowing prose
Structure as flowing paragraphs that build the argument naturally
Each paragraph should transition smoothly to the next
LIMIT each PARAGRAPH to at most 2 LONG sentences (this is very important)
Use shorter, punchier sentences within paragraphs for better readability
Create a blog post with:

1. Compelling hook that appeals to broad audience (1-2 paragraphs)

1. Style context: To capture his writing style, Tomasz uses a script that pulls from more than 2,000 of his previous blog posts stored in a LanceDB vector database. The script retrieves relevant examples and can even select material for a particular audience, like his Web3 or enterprise software readers. Retrieval gives the model references, but it does not guarantee a draft will match his voice.

He’s experimented with fine-tuning models from OpenAI and Gemma, but reproducing a real writing voice remains a challenge. Rhythm, punctuation habits, incomplete clauses, ampersands, and abrupt transitions are the tiny stylistic choices that make a voice unique, and they are exactly where AI writing still falls apart.

1. AP English Teacher review: After generating a first draft, Tomasz asks the model to grade the writing using criteria inspired by his high-school AP English class, which he credits with teaching him to love writing.

Three rounds of critique and revision

The script runs three critique-and-revision rounds. The score sometimes dips in the second round before improving in the third, a process Tomasz compares to an explore-exploit cycle. He compares all three drafts rather than assuming the final score identifies the best version.

The grading prompt evaluates the draft on several criteria:

• Letter grade and numerical score.
• Hook strength and opening sentences.
• Argument clarity.
• Evidence and examples.
• Paragraph structure.
• Conclusion strength.
• Overall engagement.

The grades move up and down across rounds, so Tomasz compares all drafts instead of trusting the final score. The AP English Teacher workflow uses a shared rubric for critique while leaving voice and final language with the writer.

AI as a first editing pass

Tomasz suggests that AI can handle the rote analysis of grammar, sentence structure, and logic, leaving teachers and writers to focus on the more human elements of argument, voice, and creative choices.

You're an experienced English teacher. Here's the letter grade numerical score, and then here are the evaluations, the hook, which, you know, argument, clarity, evidence, and examples, paragraph structure, conclusion, strength, overall engagement.

Automated feedback is immediate, which is a huge benefit for learning, but it can also be wrong or overly prescriptive. A student should use it to generate questions and revisions, not to outsource authorship or replace a teacher’s judgment.

A useful way to frame the request is to ask, "Act as a critical editor. Grade this draft against the rubric, cite the sentence behind each criticism, and explain what you would revise." The writer remains responsible for the ideas, evidence, and final language.

Two more ideas for your own workflows

Tomasz also discussed how AI may change small-company staffing and how he compares model outputs when a prompt underperforms.

A prediction about lean, AI-powered teams

At the time of recording, Tomasz predicted that a 30-person company could soon reach $100 million in revenue. He imagines a team with a product-minded CEO, 12 to 15 engineers, a small customer or developer-relations team, a salesperson for large contracts, and a solutions architect. This was a forecast, not a report on an existing company.

His broader argument is that small teams can use AI for prototyping, critique, testing, and internal tooling, though the ultimate effects on staffing and revenue remain to be seen.

Making two models 'duke it out'

When a result misses the target, Tomasz has two different models "duke it out" to fix the problem.

He gives each model the same context: the original input, the unwanted result, and the desired characteristics of the output. He then compares the revisions. Agreement between models isn't proof of quality, so he still checks the source and makes the final choice.

The value is in the contrast: different models expose different weaknesses in the same output. The AI duke-it-out workflow gives both models identical evidence and a shared rubric, then leaves the final choice with Tomasz.

Tomasz’s podcast processor is worth copying for anyone who needs to digest hours of audio for research and wants a personalized, low-friction interface. The 'AP English Teacher' writing workflow is a clever way to generate and critique first drafts, but it proves that a writer's unique voice, with all its quirks and rule-breaking, is still a deeply human skill. The final editorial judgment, like the final investment decision, remains firmly in your hands.