Teach AI to Write in Your Voice.
Generic AI scripts are slop; the model writes the average of everyone. This walkthrough turns your own transcripts into a voice document and a Claude skill that writes every script the way you actually speak.
Phone is optional. By adding it you agree Orgrowth may call or text you about this; message and data rates may apply, and you can opt out any time.
You'll also receive my content strategy newsletter with more free breakdowns.
Build your voice document and scriptwriter skill
Collect transcripts of your last 20 to 30 pieces of content
The whole system is only as good as the data underneath it, and the data is how you actually talk. Gather transcripts of your most recent 20 to 30 pieces of spoken content: reels, YouTube videos, or podcast clips. In the video I pulled mine through the Instagram graph API; if you want that route, my free Instagram audit guide walks through the exact setup (Meta developer app, access token, all of it) in about 12 minutes, linked below. But you do not need any of that. For a YouTube video, open it on desktop, click More under the title, then Show transcript, and copy it out. For reels or TikToks, free tools like CapCut or Descript will transcribe your clips, or download your own videos and let Claude transcribe them. For podcasts, your host usually has episode transcripts ready to export. Paste everything into one document, one transcript per section.
Only use content where you are speaking in your own words. Skip music-only reels, trending-audio clips, and anything scripted by someone else; they teach the AI a voice that is not yours.
Label your winners so real data gets baked in
Your voice document should know what performs, not just how you talk. Before you run the extraction, mark your top 5 pieces by views or listens and write the number next to each one, for example Top performer: [your views] views. Claude will use these to build the top performers section, so the patterns it finds are grounded in your numbers rather than guesses.
Run the extraction prompt
Open a fresh chat in Claude at claude.ai and pick the most capable model available to you. Paste in all of your transcripts with the winners labelled, then paste this prompt below them and send it. The output is your voice document; save it as a file called my-voice.md.
Use the Instagram Graph API connected to download my last 30 spoken Instagram Reels with my top performers labelled with their view counts. Study how I actually speak and produce a single markdown document called my-voice.md that teaches any AI to write in my voice.
The document must contain exactly these five sections:
1. Hook patterns: the recurring ways I open. Quote 5 to 8 real openers from my transcripts and name the pattern each one uses, for example a curiosity opener like "I just discovered this" or an urgency framing like "you are falling behind". Only include patterns that appear more than once across my transcripts.
2. Sentence rhythm: how my sentences actually flow. Typical sentence length, where I run long and where I punch short, how I transition between points, whether I ask questions or stack statements, and any verbal tics I lean on. Quote real examples for each observation.
3. Vocabulary signatures: the specific words and phrases I reach for that a generic AI would not naturally pick. List 10 to 20 of them, each with a real usage example from my transcripts. Then list the words and phrases I never use, so the AI knows what to avoid.
4. Structural templates: the beat-by-beat shapes my content follows from hook to close. Extract 2 to 3 repeatable templates, with each beat named and a one-line description of what I do in that beat.
5. Top performers: my highest performing pieces with the numbers exactly as I gave them to you, what each one did in its first three seconds, and the pattern you see behind the winners. Use only the numbers I provided; never estimate or invent a metric.
Rules: every claim about my voice must be backed by something that actually appears in my transcripts. Quote me; do not paraphrase me into blandness. If you do not have enough evidence for a section, say so instead of filling it with generic advice. Output the whole thing as one markdown document I can save as my-voice.md.Validate it before you build on it
Do not skip this. A voice document that merely sounds plausible will quietly write generic scripts forever. Ask Claude to write a test script using only the document, then hand it a script you actually wrote yourself and make it find the drift. Fold every fix back into my-voice.md before moving on.
Using only my-voice.md, write a 30 second script on [a topic you have already covered]. After you show it to me, I will paste a script I genuinely wrote myself on a similar topic. Compare the two line by line: where does your version match my hooks, rhythm, and vocabulary, and where does it drift into generic AI phrasing? List every drift, then update my-voice.md so the same drift cannot happen again.The honest test: put both scripts in front of someone who knows your content, names covered. If they can instantly tell which one is AI, run the comparison again and tighten the document.
Turn it into a real Claude skill with Superpowers
Now the second artifact: a proper Claude skill that reads your voice document every time it writes. We are not hand-rolling this; the Superpowers skill library ships a writing-skills skill built for exactly this job, and it drafts and tests your new skill for you. This runs in Claude Code. If you do not have Superpowers installed yet, type the command below into Claude Code first (a one-off). Then, in the folder where you saved my-voice.md, paste the prompt underneath it. Where the prompt says script structure, paste your own if you have one; otherwise tell it to use the first structural template from your voice document, which is a structure proven by your own content.
/plugin install superpowers@claude-plugins-officialRead my-voice.md
Use superpowers/writing-skills to create my scriptwriter skill.
The skill must instruct the AI to:
1. Read my-voice.md in full before writing a single line, every single time.
2. Write every script into this fixed structure: [paste your own script structure here, or write: use structural template 1 from my-voice.md].
3. Open with one of my hook patterns; never a generic opener.
4. Use my vocabulary signatures naturally, and avoid the words my voice document says I never use.
5. Match my sentence rhythm, checking the draft against the rhythm section and rewriting any line that does not sound like me.
6. Keep scripts to [30 to 45] seconds of spoken audio unless I say otherwise.
7. End with one clear call to action written in my voice.
8. After drafting, run a self-check against my-voice.md and fix every line that drifts generic before showing me anything.If the official marketplace install does not work, add the Superpowers marketplace first with /plugin marketplace add obra/superpowers-marketplace, then run /plugin install superpowers@superpowers-marketplace.
No Claude Code? Run it as a Project in plain Claude
Superpowers needs Claude Code, but you can get the same behaviour in claude.ai without it. Run the prompt above in a normal chat, skipping the superpowers line and asking Claude to output the skill as one markdown document instead. Then create a Project called Scriptwriter and upload both files to its project knowledge: my-voice.md and the skill document. Every chat inside that Project reads both automatically, so writing a script is one line, like the prompt below.
Follow my scriptwriter skill and write me a 30 second script about [the pain point or topic].Use it: one line per script from now on
With the skill installed, asking for a script in Claude Code triggers it automatically: it reads my-voice.md, writes into your fixed structure, and self-checks against your voice before showing you anything. This is the foundation everything downstream builds on, like the voice clone pipeline from the video, and your team can run it without you in the room.
Write me a 30 second script about [the pain point or topic].Troubleshooting: when the output still sounds generic
Three usual causes. One: too few transcripts. Under roughly 15, Claude fills the gaps with generic creator advice; add more of your content and rerun the extraction. Two: the transcripts are not actually your voice. If your reels were scripted by an editor or another AI, the document learns that voice instead of yours; feed it unscripted moments like podcast appearances, stories, or raw talking-head clips. Three: noisy sources. Music-heavy reels and trending-audio clips give the transcript lyrics and captions instead of your speech; cut them entirely.
The rule from the video: more data is not better if it is noisy. 20 clean transcripts of you genuinely talking beat 100 messy ones every time, and the same rule applies later when you train a voice clone.