# Companion prompt: turn a document I am not reading into a personalized podcast Hand this whole file to your AI. It is written for the AI, not for you. It goes with The Personalized Podcast Playbook by Erica Layer, Layer Advisory Services, layeradvisory.com. --- I want you to help me turn a long document I keep meaning to read into a private podcast episode that arrives in my podcast app on its own. Work through the five steps below in order. Do not skip ahead, and do not build anything until I have confirmed the script in Step 2, because the script decides whether any of this is worth having. **How I want you to work:** direct and competent. One step at a time, and wait for me at every point marked STOP. Tell me plainly when something I have asked for is a bad idea. No padding, no flattery. If you need a decision from me, ask rather than guessing and carrying on. **Before you start,** ask me for these three things, then reflect back what you heard in two lines: 1. The document itself, the whole thing rather than a summary. 2. Where in my week I would listen: the day and the activity. 3. What you have access to: which AI environment, whether you can write files, run shell commands, run something on a schedule, and publish to the web. Section 6 at the end is a working reference implementation. Adapt it to what I actually have rather than assuming my setup matches it, and tell me where you are deviating and why. --- ## Step 1. Check whether the length is really the problem Read the document properly. Then do two things. First, tell me what a 50% cut would destroy, quoting specific passages rather than describing them. Second, make the strongest case for cutting it in half, and show me what the shorter version looks like. Then give me your honest read: bloated, or is the length carrying the value? If it is bloated, say so plainly and recommend I cut it instead of building any of this. **STOP.** Wait for my decision. --- ## Step 2. Rewrite it for the ear This is the step that matters. Write a spoken script of the whole document, following these rules: - **Nothing is cut.** Every item, every piece of evidence, every caveat appears. If you are summarising, you have misread the task. The script usually comes out longer than the original, because speech needs connective tissue that a page does not. - No markdown: no headers, bullets, bold, numbered lists. Prose only. - No URLs read aloud. Say "it is linked in the original" if a source must be pointed at. - No section labels. Say what a section is about instead of naming it. - Numbers written the way a person says them. "Seventy six percent," not "76%." - Attribution in line and up front. A trailing parenthetical does not survive being spoken. - Shorter sentences than the written version, one idea each, since a listener cannot re-read. - Signpost every transition, because there are no visual breaks to orient against. - Repeat a name or number once where it carries weight. - Open on the single most important thing, in the first sentence. No preamble. - Second person if this is for me alone. - Output plain text, blank line between paragraphs, no title line, no notes to me. Before writing, look for anything in the document that already summarises the top item and use it as the opening. Save the script to a file. Report the word count and estimated spoken length (words divided by 155 equals minutes). Ask me to read it. **STOP.** When I come back with corrections, update the rules above rather than patching the script, and save them as a reusable spec file. I will run this repeatedly and the spec is the asset. --- ## Step 3. Narrate it Work out the cost before generating anything. Count **characters**, not words, multiply by the per character rate, then by 4.33 for a monthly figure if this repeats weekly. Compute this in code, not in prose, and show me the arithmetic. Render the whole script with a free option first. Do not spend money until I have heard a complete episode. See 6.1 for the free path and 6.2 for the paid one. **STOP.** I will listen to the whole thing. --- ## Step 4. Publish it as a private feed A podcast is a folder of audio files on the internet plus one RSS file listing them. No podcast host, no account, no database. See 6.3 and 6.4. Requirements: - Everything behind a long random path, so the address acts as the password. - The site root returns nothing, so no page lists what is there. - No blanket `Disallow: /` robots file. Podcast apps fetch feeds as bots and it will stop the feed being readable by anything. Use a `noindex` header instead. Verify against the artifact, not against a success message. See 6.5. --- ## Step 5. Make it run on a schedule Write a runbook first, in the order the steps happen, including what happens when the input is missing or the episode already exists. Then point a scheduled task at it. Put a cost limit in the runbook in writing so it is enforced every run: one AI call for the rewrite, no sub agents, no web searches. Stay silent when there is nothing to do. Write one line to a log every run whatever the outcome. Tell me how long a full run takes and record it, so a slow run is never later mistaken for a broken one. --- # 6. Reference implementation This is what actually works, from a build that runs weekly. Substitute equivalents where my setup differs. ## 6.1 Free narration, to prove the pipeline macOS has a built in voice and needs no key or network: ```bash say -v "Ava (Premium)" -o out.aiff -f script.txt afconvert -f mp4f -d aac -b 64000 out.aiff out.m4a ``` On Linux use `espeak-ng` or `piper` and convert with `ffmpeg -i out.wav -c:a aac -b:a 64k out.m4a`. Quality does not matter here. The point is one complete episode before any spend. ## 6.2 Paid narration with Gemini text to speech Key from Google AI Studio. Roughly $0.96 for a 35 minute episode on `gemini-3.1-flash-tts-preview`; `gemini-2.5-flash-preview-tts` is about half that and slightly lower quality. ``` POST https://generativelanguage.googleapis.com/v1beta/models/:generateContent?key= Content-Type: application/json { "contents": [{"parts": [{"text": "