The biggest pain with podcasts: the content is great, but a week later you can't even remember the guest's name. Transcribe it? A 90-minute episode runs 20-30k words-harder to digest than a novel. Drop it into an LLM and say "summarize"? You get a mushy overview where arguments, cases, and quotes all bleed together, useless for reuse. The problem is the instruction: no preprocessing, no output structure, no distinction between extraction and critique. This pack walks through six steps-preprocess, extract, cardify, act, structure, critique-each a standalone prompt you can run alone or chain into a pipeline. Pair it with Whisper or Tingwu for transcription, feed the text to a long-context model, and one episode becomes five kinds of reusable assets.
1. Transcript Preprocessing
Raw transcripts are unusable: "um," "yeah yeah," topic drift, no speaker labels. This step cleans the text-strips filler, segments by topic, labels speaker switches-giving every downstream step a clean input.
You are a podcast transcript cleaning assistant. Preprocess the following raw transcript:
1. Remove filler words and verbal tics ("um," "like," "you know," "right right"), but keep the natural conversational tone
2. Segment by topic shifts; add a short subtitle for each segment (≤8 words, summarizing the topic)
3. Label speaker switches as [Host] / [Guest Name]; infer from context if the transcript doesn't distinguish
4. Fix obvious ASR errors (homophones, proper nouns); mark uncertain ones as [verify]
5. Remove pure small talk and ad reads, but keep information-rich asides
6. Output the cleaned full text, preserving timestamps if present
Raw transcript: {{text}}2. Core Insight Extraction
Clean text is the foundation; now extract. This step splits each guest argument into claim, evidence, and case-three layers, not a mushy paragraph.
You are a podcast content analyst. Extract core insights from the following podcast text:
1. Each insight is a standalone block:
[Claim] one-sentence statement (≤20 words)
[Evidence] supporting reasons, 2-3 points
[Case] specific examples or data the guest cited, preserving key details verbatim
[Location] paragraph or timestamp
2. Distinguish the host's guiding questions from the guest's substantive points; extract only the latter
3. Label logical relationships between insights (progression / contrast / parallel)
4. Do not add content not in the source; mark uncertain items as [unverified]
5. Output 5-10 insight blocks, in order of appearance
Podcast text: {{cleaned transcript}}3. Quote Cards
Those "that's a great line" moments evaporate if you don't catch them. But a quote without context gets distorted. This step produces shareable short cards, each with its context-ready for social posts or note archives.
You are a quote extraction assistant. Pull shareable quotes from the following podcast text into cards:
1. Each card contains:
[Quote] verbatim excerpt, ≤40 words, keeping conversational punch, no rewriting
[Context] what question or topic prompted this, 1-2 sentences of background
[Speaker] who said it
[Use case] e.g. "social post," "card opener," "slide citation"
2. Selection criteria: has insight, has imagery, stands alone without context-reject "I think this is important" filler
3. Cut jokes and puns unless they carry a substantive point
4. Output 5-8 cards, ranked by impact
5. Every quote must be findable verbatim in the source text-no rewriting or polishing
Podcast text: {{cleaned transcript}}4. Action Checklist
Listening without acting equals not listening. This step turns the episode's methods, advice, and sparks into executable to-dos-not "prioritize health" fluff, but "log your first thought after waking, three times this week."
You are a behavior design coach. Convert the following podcast content into an actionable checklist:
1. Each takeaway format:
[Action] what to do specifically (verb-first, ≤12 words)
[Trigger] when to do it (e.g. "after standup," "replace scrolling with...")
[Expected outcome] what change you anticipate
[Source] which part of the guest's talk it comes from
2. Selection criteria: doable this week; reject grand "change your life" advice
3. Distinguish "explicit guest advice" from "your own inspiration"; label the latter [spark]
4. Note prerequisites or tools if any (e.g. "requires Notion," "needs 15 min")
5. Output 5-7 items, sorted by priority
Podcast text: {{cleaned transcript}}5. Knowledge Graph Nodes
Scattered notes aren't a knowledge base. This step extracts concepts as graph nodes, labels their relationships-"cognitive bias" links to "anchoring effect," "MVP" links to "lean startup." The output feeds directly into Obsidian, Logseq, or Roam.
You are a knowledge graph builder. Extract concept nodes and links from the following podcast text, in a format importable to note-taking apps:
1. Each node format:
## [[Concept Name]]
- Definition: ≤20 words, from the guest verbatim or your paraphrase (note source)
- Location: timestamp or paragraph
- Links: [[Related Concept A]] (relationship: part-of / opposes / extends / applies-to)
2. Label three relationship types:
- Hierarchical (A is part of B)
- Oppositional (A contradicts C)
- Extensional (A applies or extends B)
3. Preserve guest-coined terms; note the source of their definition
4. Output adjacency, e.g.: [[MVP]] → part-of → [[Lean Startup]] → extends → [[Continuous Delivery]]
5. Output 6-12 nodes, Markdown backlink format ([[concept]])
Podcast text: {{cleaned transcript}}6. Counter-Arguments and Open Questions
The most overlooked step: AI naturally defaults to "paraphrasing the guest," but guests misspeak, overgeneralize, and push pet theories. This step does critical review-which claims lack evidence, which conflict with mainstream views, which personal experiences got sold as universal laws.
You are a critical analysis assistant. Do a counter-review and open-question pass on the following podcast content:
1. Output three categories:
[Unsupported claim] the guest stated a conclusion without evidence or data; list the claim + label "lacks evidence"
[Anecdote overgeneralization] the guest framed a personal experience as a universal rule; label "sample N=1"
[Controversial view] conflicts with mainstream consensus or has significant opposition; briefly state the opposing side
2. Tag each with "verifiability": verifiable / needs checking / subjective judgment
3. Don't defend the guest, don't maliciously misread-just flag "this needs a question mark"
4. Output 3-6 items, sorted by misleading risk
5. If the guest acknowledged limitations or cited sources, label "self-corrected"
Podcast text: {{cleaned transcript}}Usage and Recommended Models
This pack is a pipeline, but each prompt stands alone. Full flow: transcribe with Tingwu or Whisper → step 1 to clean → steps 2-6 as needed. Just want quotes? Jump to step 3. Just want actions? Jump to step 4.
Model choice matters-transcripts run 20-30k words, so long context is key:
- Claude (200K context): most reliable for long-text comprehension; ideal for steps 2, 5, and 6 where you need a global view of the whole transcript in one shot.
- DeepSeek: solid Chinese-language understanding, cost-effective; fine for steps 1, 3, and 4 where you're extracting single points.
- Extra-long episodes (2+ hours): feed in segments, 20-30 min of transcript at a time through step 2, then merge and dedupe.
Pitfalls
- Dropping the whole transcript at once-context overflow, the model drops mid-section content. Feed by topic segment, with a one-line "this segment is about X" context cue.
- Sending sensitive audio to the cloud-commercial podcasts, internal interviews may carry non-public info. Transcribe with a local Whisper model; don't send sensitive content to cloud APIs.
- AI fabricates quotes-the sneakiest trap. The model "polishes" the guest's scattered words into a clean sentence that was never actually said. Every quote must be Ctrl+F verifiable in the source; if you can't find it, delete it.
- Extracting without critiquing-running step 2 but skipping step 6 turns your notes into the guest's echo chamber. What the guest says isn't automatically correct; step 6 is the honesty floor of knowledge management.
References
- OpenAI Whisper (GitHub)-open-source speech recognition model, supports local deployment, trained on 355K hours of multilingual audio
- Alibaba Tingwu-Alibaba Cloud speech-to-text tool with speaker diarization and auto-chaptering
- Anthropic Docs: Context Windows-Claude long-context usage guide and best practices
- DeepSeek API Docs-DeepSeek model API reference and context length specs