SMF·CASTMAGIC
Can a prompt replace Castmagic?
Audio & video editing — podcast recording, cleanup and publishing
Exhibit tracking slip
Verdict
Transcribing a finished episode locally and prompting an LLM to produce show notes, a blog draft, and pull-quotes from that transcript is a real weekend build, not a sitting one — getting the prompts to reliably find genuinely quotable moments (not just the first few sentences of each paragraph) takes real iteration. Unlike Alitu, this build does no audio cleanup at all; it assumes you already have a finished recording and repurposes the words, not the sound.
Exhibit A — The prompt
Received on31.07.2026Build a local content-repurposing tool: transcribe a finished recording, then generate show notes, a blog draft, and social quotes from the transcript with an LLM. Stack: Python, whisper.cpp for local transcription, Anthropic API for content generation, SQLite for storing transcripts and outputs. Do not build any audio cleanup or editing — that's a different tool's job.
Core loop: the user uploads a finished audio file (already mixed and mastered elsewhere). Transcribe it locally with whisper.cpp, then send the transcript to Claude with separate prompts for: episode show notes (a summary plus timestamped topic list), a rough blog-post draft in the speaker's own words rather than a generic rewrite, chapter markers derived from topic shifts in the transcript, and 5–8 short, quotable excerpts flagged with their timestamp for social posts. Show every generated piece next to the source transcript passage it came from, so the user can verify it against what was actually said before publishing anything.
Requires an Anthropic API key — the whole point of this tool is the LLM-driven repurposing step, so unlike a pure transcription tool this genuinely needs one. Without a key, transcription still works and produces a plain transcript, but none of the derived content generates.
Do not build: audio cleanup, mastering, or episode assembly (that's Alitu's job, not this one), direct publishing integrations to podcast hosts or CMSs, or a prompt-template marketplace — ship one well-tuned prompt per output type rather than a library of them.Opening prefills the prompt — press enter to run it.
Exhibit B — What you lose
- B.1 any audio cleanup or mastering — this repurposes words from an already-finished recording, it does not touch the sound
- B.2 a large library of tested prompt templates tuned per content format
- B.3 integrations that publish directly to a podcast host, CMS, or social scheduler
- B.4 hosted processing, so a long episode doesn't tie up your own machine
- B.5 team workflows for reviewing generated content before it ships
Prior art
Exhibit C — Why people still pay: audio infrastructure, distribution, and production polish
People pay for Castmagic because getting an LLM to reliably pull the three or four moments in an hour-long conversation actually worth quoting — not just the loudest or most recent ones — takes prompt engineering tuned across a huge volume of real transcripts, plus a direct publishing pipeline into the tools creators already use. A personal build gets a working first draft; it doesn't get that tuning or the publish step.
Questions
Can I import my existing Castmagic transcripts and outputs?
Castmagic lets you export transcripts and generated notes as text, which paste in directly. Regenerating the derived content (show notes, quotes) from this build's own prompts will produce different results than Castmagic's tuned versions.
Will it work on my phone?
Reviewing generated show notes and quotes works fine in a phone browser. Uploading and transcribing a large audio file is more reliable from a desktop with a stable connection.
What does it cost to run?
Transcription with whisper.cpp is free (local compute only). Generating show notes, a blog draft, and quotes for one episode costs roughly $0.10–0.30 in Claude API usage, depending on episode length.
What's the one thing that doesn't survive the rebuild?
Consistently good quote selection. A well-tuned prompt finds real, genuinely quotable moments; a first-pass prompt here tends to pull whatever's near the start of a sentence rather than what's actually striking, and that gap only closes with real iteration against your own transcripts.
Related tools
Receipt