Which tool to choose
They're not interchangeable. Choose by what you have to produce, not by trend.
You need to write the text you'll read or perform yourself (reel, YouTube video, voiceover). Use a text chatbot: ChatGPT, Claude, or Gemini. They're the most flexible for rewriting, shortening, and adapting the tone. Watch out for ChatGPT's free plan: free users get about 10 messages every 5 hours on the default model, then drop to a faster but less capable mini model. In a writing session with many revisions the limit runs out fast: alternate two free tools (Claude and Gemini) or concentrate everything into a few well-made prompts.
You need to produce a podcast starting from already-written material (an article, your notes, a PDF, some research). Use Google's NotebookLM. It's free with a Google account, with no paid plans or limits on basic usage. It generates a conversation between two AI voices starting from your sources, connecting the loaded topics in podcast style.
You want a structured narrative script without writing prompts. Tools like Squibler offer dedicated editors, but they don't let you set length or structure in detail: they produce a draft to touch up, fine for getting started and less so for fine control. For most people a generic chatbot plus the right prompt gives more control at zero cost.
How to do it
The principle that holds everything together: the AI writes well only if you give it precise context. "Write me a script for a reel" produces generic fluff. You have to tell it format, length, audience, goal, and tone.
For a script to perform or read (reel, video, narration)
- Open ChatGPT, Claude, or Gemini.
- Paste this prompt, replacing only the parts in quotes with your data:
You are a screenwriter who's an expert in short-form content.
Write me the script for a 30-second Instagram reel.
Topic: "how to save on your electricity bill"
Audience: "Italian families who don't understand anything about energy tariffs"
Goal: "get them to follow me because I explain practical things simply"
Tone: direct, friendly, zero technical jargon
Required structure:
- Hook in the first 3 seconds: a sentence that immediately frames
the problem and who the video is for
- Development: a single idea, taken all the way
- Closing: a coherent call to action (follow, comment, or save)
Constraints:
- Maximum 55 words total, because I'll perform it on camera
- Short, spoken sentences, the way I say things out loud
- Indicate in parentheses when a sentence should also be shown as
on-screen text
Give me 3 versions of the hook to choose from.
- Read the draft aloud. If you trip over a sentence, it's too written: ask "rewrite this sentence the way I'd say it speaking."
- To shorten: "cut this to 40 words keeping only the central message."
Feedback: if what you read sounds like a press release, you gave little context about the tone. Add an example of how you talk ("write as if you were explaining it to a friend at the bar") and regenerate.
The word limit isn't arbitrary: for a 15-second reel don't go over 40-50 words if you're speaking on camera; if you use only on-screen text, aim for 5-7 words per screen, so it stays readable. Adjust the figure to the actual length.
The three-act structure holds up any short format: Act 1, the hook, the promise that answers "why should I stay?"; Act 2, the value, the central concept or the solution to a specific problem; Act 3, the closing, the coherent call to action.
For a podcast from written sources (NotebookLM)
- Go to notebooklm.google.com and sign in with your Google account.
- Create a notebook and upload the sources: it handles up to 50 sources among PDFs, website links, YouTube videos, and pasted text.
- Open the Audio Overview section in the Studio panel and choose the format (Deep Dive, Brief, Critique, or Debate). Where available, set focus areas, technical level, tone, and length.
- Start the generation: the audio is ready in a few minutes, usually between 6 and 15.
Feedback: the podcast should reflect your sources, not fire off generic opinions. The discussions are neither complete nor objective: they're a reflection of the material you load. If it wanders, narrow the sources or use the customization field to indicate the points to cover.
For the customization instructions, keep them under about 1,500 characters: overly long entries can cause the model to prioritize some instructions unpredictably.
A concrete example
Marco runs a small tax-consulting practice and wants a reel on the topic "flat-rate tax regime." He opens Claude (he's already used up ChatGPT's free messages from the morning) and pastes the prompt above, adapted: topic "flat-rate regime yes or no", audience "freelancers in their first year of a VAT number", goal "show that I explain taxes without terror", tone "reassuring and practical".
The AI returns three hooks. Marco chooses: "You open a VAT number and everyone tells you 'flat-rate.' But for you it might be a mistake." The development explains in two sentences when the flat-rate regime isn't worth it. The closing: "Save this before you go to your accountant."
Marco reads it aloud: the phrase "revenue threshold" trips him up. He asks "rewrite without accountant terms." It becomes "if you take in more than X per year." Result: 48 words, recorded in one take, subtitles included. From idea to finished script: 8 minutes.
It works because it respects the golden rule: a single idea per reel. Marco was tempted to also cram in VAT and social-security contributions, but he moved them into two separate reels.
When it doesn't work (and how to fix it)
If the AI invents numbers, data, or quotes
Chatbots generate plausible, unverified text. If your script cites "73% of Italians", don't trust it: that number could be made up. Fix: write in the prompt "don't include statistics or percentages, I'll leave the gaps for data I'll verify." For a podcast on certain sources use NotebookLM, which works only on the documents you upload; even there, though, small inaccuracies can appear in interpreting details, so double-check the very technical passages.
If the script sounds fake, "AI-ish"
The typical flaw: sentences that are too perfect, big words, flat rhythm. Fix: paste 3-4 real sentences of yours into the prompt (an old post, a transcribed voice note) and write "imitate this way of speaking." The AI calibrates the tone to yours, not to its neutral textbook register.
If you run out of free messages mid-job
On ChatGPT's free plan, after the limit you snap to the mini model and quality drops. Immediate fix: switch to Claude or Gemini (free, separate account) and continue there by pasting in the draft you already have. Don't wait for the reset: the limits reset on a rolling window, not at midnight, so the countdown starts from your first message.
If the NotebookLM podcast comes out in English or monotone
Audio Overview supports over 80 languages for audio generation; the interactive "join the conversation" mode started in English and support is expanding. To change the language: NotebookLM uses by default the one preferred in your Google account settings, so set Italian there. If the voices are flat, choose the "Debate" or "Deep Dive" format instead of "Brief".
If the AI gives you a long, dense script
The most common and most damaging problem. The frequent mistake is wanting to say too much in a few seconds: the reels that work develop a single concept, taken all the way. Fix: ask "does this script contain more than one idea? Split it into separate scripts, one per concept." You turn a confused video into a series.
A tip from someone who actually uses it
Write the hook last, not first. It seems counterintuitive, but when you already have the central message clear in front of you, generating 5-10 opening variants becomes easy and targeted. The hook decides everything: according to Meta's internal data users choose whether to stay or scroll in about 1.7 seconds, and if the reel doesn't grab attention in the first 3 seconds the subsequent distribution suffers.
A mindset shift: for years people talked obsessively about the hook, but today the point isn't to shock at all costs, it's to orient the viewer immediately. In the first 2-3 seconds it must be clear what you're talking about and who the video is for. When you ask the AI for hooks, specify "clear and orienting, not clickbait": clickbait openings cause immediate skips, lower the initial retention, and compromise long-term distribution.
Frequently asked questions
How long should the script for a reel be?
It depends on the video's length, but count the words, not the lines. For speaking on camera, about 2-3 words per second: a 20-second reel stays under 50-55 words. Today creators typically have reels up to 90 seconds as the standard, with tests up to 180, but the 15-30 second range remains the feed's fast lane. Short and clear beats long and dense.
Can I do everything for free or do I have to pay?
For the script alone, yes: ChatGPT, Claude, and Gemini have free plans sufficient to write and revise, with the message limit to manage. For the podcast, NotebookLM is free with no real usage caps. You pay when you want to automate exporting videos: with tools like PlayPlay the script generation is free, but exporting without a watermark requires a paid plan.
Should I write the script first or improvise on the fly?
Always the script first. Even if you then improvise, having the structure in mind keeps you from the confused video that the algorithm punishes. The script isn't a cage: it's the rail that keeps you on a single idea.
What language do I write the prompt in?
In Italian. Current models work great in Italian and return a script already in the right register. Writing the prompt in English to "get better results" is an outdated myth: you'd only introduce loan-translations and unnatural phrasings to fix afterward.
Doesn't the AI steal my voice and personality?
The problem only arises if you publish the raw output. The AI is a draft accelerator, not the final author. The value of your content lies in the point of view, in the examples drawn from your experience, and in the way you say things: parts you provide, in the prompt and in the revision. If you paste and publish without touching anything, you'll sound like everyone else doing the same. If you use the AI to get past the blank page and then put your voice into it, you get the best of both.