There are two ways in, and picking the wrong one is where people waste a weekend.
Two of the nine really want your own voice, builds 2 and 7. Three of them are deliberately better with a library voice. Do this once if you want it, then never again.
Builds 1, 2, 3 and 9 talk to Fish Audio as your account, not through the website, so they need a key. Do this before you open any of the prompts.
.env holding one line, FISH_API_KEY=your-key-here, and a file called .gitignore containing .env. Never paste a key into a chat window, a web page or a screenshot.{{system.caller_number}}. Anything you assumed would just arrive will not.localhost and 127.0.0.1 count as different origins. List both, exactly, or the widget silently refuses to connect.Answers your business phone around the clock in a natural voice, answers questions about what you do, and books appointments straight into your calendar. The thing it really sells is the 7pm call you would otherwise miss.
You are setting up an AI phone receptionist for my business. Read https://docs.fish.audio/agents/overview before you start.
First, look at what I already have: tell me which of these exist and which I need to create, without guessing.
- a Fish Audio account with Agents enabled
- a phone number I can route
- a calendar with a booking link
Then ask me these, one at a time, with a default I can accept:
1. What does my business actually do, in one sentence? Default: ask me to paste my website's about line.
2. What are the three questions callers ask most? Default: opening hours, price range, and whether you take walk-ins.
3. What should it never promise? Default: no discounts, no guaranteed times, no medical or legal advice.
4. What happens when it cannot help? Default: take a message and say someone will call back today.
5. Should calls be recorded? Default: OFF, and if ON the agent must say so in its first sentence. Recording defaults to on in Fish Audio, so this is a setting to change, not a box to leave alone.
6. Which calendar or booking tool, and does my plan on it actually include API access? Default: a free Cal.com account, but check its free tier before we build anything, and if it does not include API access, name one whose free tier does.
Then propose the agent's first message and system prompt, and wait for my yes before building.
Build it in these steps, stopping after each:
(1) create the agent and set the voice
(2) write the system prompt from my answers
(3) add my FAQs as a knowledge base file
(4) connect the calendar so it can actually book, as a webhook tool that checks availability and writes the booking, passing the caller's name and the time as declared tool arguments, served behind a real public address because Fish Audio rejects localhost
(5) publish, and tell me the version
Test with me first. Pass means: it answers in under two rings, gets a made-up question right, refuses something it should refuse, and books a real slot I can see.
Tell me exactly what this costs per call, and make its first sentence say it is an AI assistant. It never claims to be human.Calls every new lead within a minute of them filling in your form, asks three qualifying questions in your own cloned voice, and writes the summary back to your CRM. Speed is the entire product here.
You are building a speed-to-lead caller. Read https://docs.fish.audio/agents/telephony/outbound-calls first.
Look at what I have: tell me which exist and which to create.
- where my leads arrive now (form, ad lead form, booking tool or CRM)
- whether that tool can send a webhook
- a Fish Audio voice clone of me
Ask me one at a time, each with a default:
1. What three questions qualify a lead for me? Default: what they need, when they need it, and whether there is a budget.
2. What counts as qualified? Default: they answered all three and the timeline is inside 90 days.
3. What should happen to an unqualified lead? Default: polite close, notes written, no follow-up call.
4. When may it call? Default: 9am to 6pm in the lead's own timezone, never weekends.
5. Consent: has the lead agreed to be contacted by phone? Default: assume NO, and refuse to build until there is a ticked consent box, because calling without it is illegal in most places.
Propose the call script and wait for my yes.
Then: (1) webhook in, (2) format the number to E.164, (3) check consent and de-duplicate, (4) place the call, (5) write the summary back.
Test with me first. Pass means: I submit a test form with my own number, my phone rings inside 60 seconds, it asks all three questions, and a readable summary appears on the contact.
Tell me the cost per call and per minute, and make it say it is an AI assistant in the first sentence.Answers customer questions from your own FAQs and policies day or night, and hands the hard ones to a person with a summary of what was already said, so nobody repeats themselves.
You are building an AI support line. Read https://docs.fish.audio/agents/build/knowledge-base and https://docs.fish.audio/agents/telephony/transfers first.
Look at what I have: which of my policies already exist as text, and which are only on my website.
Ask me one at a time, with defaults:
1. Which three topics do customers call about most? Default: order status, returns, delivery times.
2. What must it never answer on its own? Default: anything about a refund over US$200, anything legal, anything medical.
3. Who does it transfer to, and when? Default: a human on anything it cannot source from the documents.
4. What does it say while transferring? Default: a one-line summary so the person is not starting cold.
5. How current are the documents? Default: assume stale, and tell me which ones to re-check before go-live.
6. Should calls be recorded? Recording defaults to on in Fish Audio. Explain in two lines where I stand on consent in my own country, then ask me to choose. Default: off until I have checked.
Propose the answer style and wait for my yes.
Then: (1) load the documents, (2) write the system prompt with a hard rule that it only answers from them, (3) set up warm transfer, (4) add the summary-on-transfer step, (5) publish.
Test with me first. Pass means: it answers a real policy question correctly, it refuses a question that is not in the documents instead of inventing an answer, and a transfer arrives with a usable summary.
The single most important rule: if it is not in my documents, it says it will find out, and transfers. Its first sentence always says it is an AI assistant.Writes a batch of UGC-style ad scripts for your product and voices each one differently over the same footage, so you can test four angles instead of betting the budget on one.
You are writing and voicing UGC-style ad variations. Read https://fish.audio/plan/ first and tell me which plan I need, because commercial use is not on the free plan.
Look at what I have: the footage length, whether it has usable audio already, whether I have a cloned voice or need a library voice, and whether ffmpeg is installed on this machine.
Ask me one at a time, with defaults:
1. What does the product actually do for someone, in one sentence? Default: ask me to paste my product page's first line.
2. Who is this ad for? Default: the person who already has the problem and is searching for a fix.
3. What is the one objection that kills the sale? Default: price.
4. Where is it running? Default: Instagram Reels, sound on, first three seconds decide everything.
5. Should the voice be mine or a library voice? Default: a library voice, because an ad in your own voice sets an expectation you have to keep.
Propose four angles, one line each, and wait for my yes before writing full scripts. The four should be: a problem opener, a numbered list, a curiosity opener, and an honest review.
Then: (1) write the four scripts at 15 seconds each, (2) pick a different voice for each, (3) generate the audio, (4) transcribe each finished voiceover back with timestamps to build the subtitle file, then use ffmpeg to cut the footage to each voiceover's length, lay the voice over it and burn in the captions. Ask me before you install anything. (5) export four files named by angle.
Test with me first. Pass means: four video files exist, each plays its own voice in sync with its own captions, and the four openings are genuinely different claims rather than four rewordings of one.
If I need timestamps for captions, transcribe-1-pro is US$0.36 per audio hour, so tell me that cost too.
Never invent a statistic or a customer quote for an ad.Plays the hiring manager for the specific job you are applying for, runs the interview out loud, and scores you afterwards on the things that actually lose offers.
You are running a mock interview. Read https://docs.fish.audio/agents/deploy/widget and https://docs.fish.audio/agents/monitor/post-call-analysis first.
I will paste the job ad. Before you ask me anything, read it and tell me the three things this employer clearly cares about most, so I can correct you if you have misread it.
Then ask me one at a time, with defaults:
1. What level is this role? Default: infer it from the ad and tell me what you inferred.
2. Which two of my projects should I be ready to talk about? Default: ask me to name them.
3. How long should the interview run? Default: ten minutes, about six questions.
4. How hard should it push? Default: polite but unsatisfied, it follows up once on every vague answer.
5. What am I most worried about? Default: ask me, and make it ask that question twice in different words.
Then run the interview out loud. One question at a time. Never show me the next question while I am answering.
Afterwards, score me out of ten on each of: clear structure, specific numbers, answer length, and the questions I asked them. For each, quote something I actually said and say what would have been better.
End with the single thing that would have lost me the offer. One thing, not a list.
Pass means: I speak six answers out loud, it follows up at least once on a vague one, and a scorecard appears naming one thing to fix. Before we start, tell me in one line whether I end up with a published Fish Audio agent on a page I serve, or a voice conversation inside this tool, so I know what I am left holding.A patient tutor you have real spoken conversations with, which corrects your mistakes as you go and keeps the conversation at your level instead of dropping you in the deep end.
You are my spoken language tutor. Read https://docs.fish.audio/agents/build/voice-language first and tell me which of the speaking languages is closest to what I want.
Ask me one at a time, with defaults:
1. Which language, and which variety? Default: ask, because Latin American and European Spanish are different enough to matter.
2. What is my level, honestly? Default: assume I can order food and not much else, and adjust up if I prove otherwise.
3. What do I actually want to do in this language? Default: hold a conversation with a stranger without panicking.
4. How should corrections work? Default: correct in passing, keep talking, never stop the conversation to teach a rule.
5. How long per session? Default: ten minutes, because concentration goes after that.
6. Should it switch to English when I am completely stuck? Default: no, target language only. If yes, turn multilingual recognition on, because with it off the agent listens only for the speaking language and will not react to my English at all. Never turn strict language on for a tutor.
Then have an actual conversation with me. Rules you must hold:
- Speak only in the target language unless I am clearly lost
- When I make a mistake, say the correct version once, naturally, then carry on
- Never correct more than one thing per turn, even if I make three mistakes
- Keep your sentences at my level, not at yours
At the end, give me the three mistakes I made more than once, each with the rule in one line of plain English. Nothing else, no encouragement paragraph.
Pass means: I hold five minutes in the target language, it corrects me mid-flow without stopping the conversation, and I come out with three mistakes I repeat. Tell me first whether this ends up as a published agent on a page I serve, or a voice conversation inside this tool.Turns a book, a course or a long guide into finished audio in one consistent voice, so chapter 18 sounds like the same person on the same day as chapter 1.
You are producing an audiobook. Read https://docs.fish.audio/features/voice-cloning first.
Look at what I have: how many chapters, roughly how many words in total, and whether the text has headings I should not read aloud.
Ask me one at a time, with defaults:
1. Whose voice? Default: mine, cloned once and reused, because a library voice is fine for one chapter and obvious across twenty.
2. Which words will it get wrong? Default: ask me for names, brands and any technical term, and build a pronunciation dictionary before generating anything.
3. How should chapter openings sound? Default: a short pause, the chapter title, a longer pause, then the text.
4. What do I do with headings, footnotes and captions? Default: skip them, and tell me which ones you skipped.
5. Where is this going? Default: ask, because a sold audiobook needs a paid plan and a personal one does not.
Propose the pronunciation dictionary and wait for my yes, because fixing it after twenty chapters means regenerating twenty chapters.
Then: (1) clean the text per chapter, (2) generate chapter one only, (3) stop and let me listen, (4) generate the rest once I approve, (5) give me the files named and in order.
Pass means: I play the end of one chapter into the start of the next and cannot hear a change.
Tell me the total cost before you generate anything past chapter one. For reference, s2.1-pro is US$15 per million UTF-8 bytes, which is roughly 180,000 English words or about 12 hours of speech, so price my manuscript against that.Writes and voices a whole week of videos without you being on camera once, so the channel keeps posting while you are doing something else entirely.
You are batching a week of faceless video narration in the Fish Audio web app. Read https://fish.audio/plan/ first and tell me which plan I need, because a monetised channel is commercial use and that is not on the free plan, and the free tier caps each generation at 500 characters.
Ask me one at a time, with defaults:
1. What is the one topic for this week? Default: ask, and refuse to proceed with seven unrelated topics, because the point is one subject from seven angles.
2. Who is watching? Default: someone who has the problem but has never heard the jargon.
3. How long is each video? Default: 60 seconds, about 160 words.
4. What is the through-line? Default: every video ends on the same idea said differently, so the week adds up to something.
5. Whose voice? Default: a library voice, chosen once and locked, because seven files in seven different voices is not a channel.
Propose the seven angles as one line each and wait for my yes. Reject your own first list if two angles are really the same angle.
Then: (1) write all seven scripts, (2) read them back to me as a list so I can cut one, (3) generate all seven in one voice, (4) name the files by angle, (5) tell me the total cost.
Pass means: I can play any two back to back and they sound like the same person on the same day, and no two scripts open with the same sentence shape.
Do not write a hook you cannot deliver in the body. That is the one thing that kills a faceless channel.Reads your notes, your calendar and what is actually due back to you as a five-minute audio file while you get ready, instead of you scrolling three apps before you have had coffee.
You are building a spoken morning brief. Read https://docs.fish.audio/overview/mcp first, and tell me the exact steps to connect the Fish Audio MCP server for the tool I am using.
Look at what I have: which of my notes, calendar and task list you can actually read, and say plainly which ones you cannot.
Ask me one at a time, with defaults:
1. What time should it land? Default: 6:30am, so it is there before I pick up my phone.
2. What goes in it? Default: what is on today, what moved since yesterday, and the one thing that matters most.
3. How long? Default: about 700 words, which is four to five minutes read aloud.
4. What should it never include? Default: anything from a folder I name as private, and never read out a password or a number from a bank file.
5. What tone? Default: plain and warm, like a person telling me, not a list read aloud.
6. Which voice? Default: a library voice, or my saved voice ID if I have cloned one. Ask me for the ID rather than guessing it.
Propose one sample brief from today's real data and wait for my yes before scheduling anything.
Then: (1) the read step, (2) the writing step, (3) the voice step, (4) save to a synced folder, (5) schedule it daily.
Pass means: tomorrow's file exists before I wake up, is under six minutes, and tells me one thing I had forgotten.
Say out loud what leaves my machine, because this reads my actual notes.