Build the 3 Things
From My Drama 3 Video

const SECRET = 'change-me-to-a-long-random-word';

function doPost(e) {
  const data = JSON.parse(e.postData.contents);
  if (data.secret !== SECRET) {
    return ContentService.createTextOutput('unauthorized');
  }
  const sheet = SpreadsheetApp.getActiveSpreadsheet().getSheetByName('Bookings');
  sheet.appendRow([
    new Date(),
    data.name || '',
    data.phone || data.caller_number || '',
    data.requested_time || '',
    data.service || '',
    data.notes || '',
  ]);
  return ContentService
    .createTextOutput(JSON.stringify({ saved: true }))
    .setMimeType(ContentService.MimeType.JSON);
}

You're going to build me an AI phone receptionist inside Fish Audio using their Voice Agents API. Use the Fish Audio skill I installed, and read https://docs.fish.audio/llms.txt for the current endpoints before you call anything.

STEP 1. SETUP
Ask me for my Fish Audio API key and save it in a .env file as FISH_API_KEY. Never print it back to me.

STEP 2. INTERVIEW ME
Ask me these one at a time and wait for each answer:
1. Business name and what we do
2. Hours, including holidays
3. Every service with prices or price ranges
4. Address, parking, and how to find us
5. Service area, if we travel to customers
6. How booking works today (appointment length, how far out we book)
7. What to say to sales calls and solicitors
8. The 5 to 10 questions callers ask most, with the answers
9. What to do when it doesn't know the answer
10. The name and personality I want the receptionist to have
11. My area code (for the phone number)
12. My Google Sheets web app URL and the secret word from my Apps Script

STEP 3. WRITE THE BRAIN
Write a system prompt under 2,000 tokens. It's a voice agent, so: one or two spoken sentences per reply, no lists, no markdown, never invent prices or availability, and when unsure take a message (name, callback number, reason, then read it back). The greeting says "this call may be recorded."
Booking rule: when a caller wants an appointment, collect their name, phone number, the day and time they want, and the service. Repeat it back. Once they confirm, call the log_booking tool, then tell them they're on the schedule and the team will confirm shortly.
Then write a knowledge doc as plain .txt with every fact from the interview.
Show me both and wait for my OK.

STEP 4. BUILD IT IN FISH AUDIO
1. Upload the knowledge doc as a knowledge source.
2. Show me 3 warm, natural-sounding voices from the Fish Audio library and let me pick one.
3. Create the agent with the system prompt, the voice I picked, and a fixed first message that uses the business name and the receptionist's name. Attach the knowledge source.
4. Create a webhook tool named log_booking. Arguments: name, phone, requested_time, service, notes. Method POST to my Sheets web app URL, content type application/json, execution mode fire_and_forget. Body template:
{"secret": "MY_SECRET", "name": "{{name}}", "phone": "{{phone}}", "requested_time": "{{requested_time}}", "service": "{{service}}", "notes": "{{notes}}", "caller_number": "{{system.caller_number}}"}
Attach it to the agent.
5. Publish the agent.

STEP 5. PHONE NUMBER
Search available local numbers in my area code. Show me 3 options and the monthly price. Do NOT buy anything until I say yes. When I pick one, buy it and bind it to the agent.

STEP 6. HAND OFF
Give me the phone number, a link to the agent in the Fish Audio console, and a 5-call test script (a question that's in the doc, one that isn't, a booking, a solicitor, and a caller who interrupts). Remind me that callers only hear the published version, so I have to publish again after every edit.

Write my morning brief, then turn it into audio.

1. Pull what matters from the last 24 hours:
- Gmail: anything from a real person that needs a reply or a decision. Skip newsletters, receipts, and promos.
- Google Calendar: everything on my calendar today, with times.
- My notes: [NOTION PAGE / GOOGLE DOC / FOLDER NAME] for open to-dos and anything I flagged.

2. Write it as a script for the ear, 2 to 3 minutes out loud (about 350 to 450 words):
- Open with "Good morning" and the single most important thing today.
- Then today's schedule, then the emails that need me (who, what they want, what I should do), then my top 3 to-dos.
- Close with one line to start the day.
- Short sentences. No bullets, no headings, no links, no emoji. Write times the way you'd say them ("two thirty", not "2:30").
- Add delivery directions in square brackets where they help, like [warm, upbeat morning radio host] at the top and [pause] between sections.

3. Use the Fish Audio connector to read the script as speech in my cloned voice, "[YOUR VOICE NAME]". Use the Drama 3 model if the connector lets you pick one.

4. Give me the audio link first, then the written script underneath so I can skim it.

I want a personal podcast waiting on my phone every morning: Claude writes a brief from my email, calendar, and notes, and Fish Audio reads it in my cloned voice. Build it on this computer. Use the Fish Audio skill I installed and read https://docs.fish.audio/llms.txt before writing any Fish Audio code.

Ask me first, one at a time:
1. Where my email and calendar live (Gmail, Outlook, other) and where my notes live (a folder of .md files, Notion, Google Docs, Apple Notes)
2. My Fish Audio API key (save it in .env as FISH_API_KEY, never print it)
3. The name of my cloned voice in Fish Audio (look up its voice ID for me)
4. What time I want it ready, and which folder the MP3 should land in so it shows up on my phone (an iCloud Drive or Google Drive folder works)

Then build:
- A step that gathers the last 24 hours of email that needs me (skip newsletters, receipts, promos), today's calendar, and open to-dos from my notes.
- A step where Claude writes the brief (headless claude -p is fine): 350 to 450 words, written for the ear, short sentences, no lists or links, times written the way you'd say them. Open with the most important thing today, close with one line to start the day. Use Drama 3 directions in square brackets, like [warm, upbeat morning radio host] at the top and [pause] between sections.
- Text to speech through the Fish Audio API with my voice ID and the model header set to drama-3. If that header is missing the request quietly falls back to an older model, so double check it. Save it as YYYY-MM-DD-morning-brief.mp3 in my chosen folder, with a .txt copy of the script next to it.
- A daily schedule that runs the whole thing at my chosen time (use whatever is most reliable on my operating system) + a log file so I can see if a run failed.

Run it once now so I can listen. Then tell me how to change the time, the length, or the sections.

Help me make a short animated video (30 to 45 seconds) with AI images, AI video, and a Fish Audio Drama 3 voiceover.

First ask me 3 things, one at a time:
1. The idea (a funny scene, a mini story, an explainer with characters, a fake movie trailer, anything)
2. The look (Pixar-style 3D, anime, claymation, 90s cartoon, storybook watercolor)
3. Vertical 9:16 for Reels and TikTok, or wide 16:9

Then give me 4 things:

1. CHARACTER SHEET
For each character (3 max): a name, a one-line personality, and one detailed image prompt for a clean full-body character reference on a plain background, in the chosen style. I'll generate these first and reuse them as references so the characters stay consistent from shot to shot.

2. SHOT LIST (6 to 8 shots, about 5 seconds each)
For each shot: what happens, an IMAGE PROMPT for the starting frame (characters, setting, camera angle, lighting, style), and a MOTION PROMPT for animating it (one action + one camera move, like "she slams the laptop shut, slow push-in"). One action per shot. AI video falls apart when you ask for three things at once.

3. VOICEOVER SCRIPT FOR DRAMA 3
Write it as a scene with speaker names. Put a direction in square brackets before each line, the way you'd brief a voice actor: [barely holding it together, forcing a smile] or [whispering, way too excited]. Add sounds and pauses where they happen: [sigh], [small laugh], [pause]. Wrap the one word that has to land in <emphasis></emphasis>. Directions apply to everything after them, so put each one right before the words it's for. Label every line with its shot number.

4. VOICE CASTING
For each character, describe the voice I should search for in the Fish Audio voice library (age, accent, energy).

Generate a multi-speaker voiceover with Fish Audio Drama 3 in one pass. Use the Fish Audio skill and read https://docs.fish.audio/features/text-to-speech/drama-3 first.

My API key is in .env as FISH_API_KEY (ask me for it if it isn't there).

Here's my script and the voice I picked for each speaker:
[PASTE THE VOICEOVER SCRIPT CLAUDE WROTE]
[SPEAKER NAME = voice name or voice ID from the Fish Audio library]

Do this:
1. Look up the voice ID for any voice I gave you by name.
2. Turn the script into one text string: swap each speaker name for a speaker tag (<|speaker:0|> for the first voice, <|speaker:1|> for the second, and so on) and keep every [direction] and paired tag exactly as written.
3. Call POST /v1/tts with the model header set to drama-3 and reference_id as an array of the voice IDs in speaker order. Save it as voiceover-take-1.mp3.
4. Drama 3 is still a preview model, so takes vary. Make 3 takes (take-1, take-2, take-3) so I can pick the best one.
5. Tell me which shot each line lands on so I can line it up in my editor.