Voice Kit is a skill for one useful job: take a piece of writing you stand behind, pair it with the thing you need to write today, and make a first draft without inventing a fake version of your voice.
Paste the one-time prompt into the AI you already use. Then give it one past email, post, transcript, or paragraph—and the job in front of you. See the exchange first.
Current GitHub version
Voice KitChecking GitHub…
Checking the published version and update notes.
Use it once first
Start with the work already in front of you. The first run should leave you with a draft you can edit and a clear signal about whether Voice Kit reduces the part of writing you actually dislike.
Use the button above. You do not need to install anything for the first run.
“Here is an email I wrote. Write a short follow-up to this founder about the meeting we just had.”
Tell it what sounds unlike you, what fact is wrong, or what it missed. That is the useful feedback.
See the exchange
This is a synthetic example. It shows the shape of the work, not a performance result or a promise that every draft will land on the first try.
Sample: “The hard part of using AI well is usually not getting words on the page. It is deciding which real detail belongs in the first sentence.”
Assignment: “Write a short follow-up after meeting a founder. Keep it direct. Ask about one workflow they are trying to make useful.”
Draft: “Good meeting you. The question I keep coming back to is which workflow needs to become useful first. If you have one you are trying to improve, send it over. I would be interested in seeing where the real friction is.”
The output is useful because you can point to the source, see the choice, and correct the line that is off. You are not being asked to trust a personality label.
What the skill keeps separate
Voice Kit is the reusable instruction. Your Corpus holds the real writing and corrections you choose to keep. Today’s assignment gives both of them a specific job to do.
The writing you share, plus corrections you choose to save. It stays close to the source instead of getting flattened into a personality label.
“The hard part of using AI well is usually not getting words on the page. It is deciding which real detail belongs in the first sentence.”
“If you have one workflow you are trying to improve, send it over. I would be interested in seeing where the friction is.”
Use only material you own or are allowed to use. The skill asks before creating or updating a saved Corpus file.
The Skill your AI uses to read the source, draft from today’s assignment, show the pattern notes, and take your correction seriously.
Your direct correction wins over a pattern the AI inferred, whether you save it for later or use it only in this session.
Illustrative working context. In a skills-capable AI, Voice Kit is the reusable instruction and the Corpus is private context you control. In a normal chat, use the current sample and feedback for that session. The skill never creates or updates a saved Corpus file until you approve it.
Optional details
Optional setup
The one-time prompt above is enough for a first run. Install Voice Kit when you want the Skill available for future writing work. You still provide the current sample and assignment every time, and you choose whether any private Corpus is saved or updated.
The hero button copies this whole instruction. Read it when you want to see the checks, feedback loop, and limits behind the draft.
# Voice Kit: make AI draft like me, not like AI
Give me a draft in my actual voice from whatever real samples I share, a YouTube link, a doc, a pasted email, even one paragraph. Don't run an interview before drafting.
The one idea that matters: a voice kit built only from abstracted rules ("be direct," "avoid jargon") produces AI-sounding output, because a model given a rule like that invents new phrasing that satisfies the rule instead of recalling my actual words. The fix: keep my real sample text on hand, and check anything punchy-sounding against it before using it.
How to work:
1. Take whatever I give you, immediately. One sample is enough to start. If more would sharpen it, mention that once after the first draft, not before.
2. Draft from my current request right away if it already states the audience, channel, and topic. Only ask a direct question when something you actually need for the draft is missing.
Two ways I can give you a sample, either is fine:
- I paste or link it: a YouTube URL, a blog post, a pasted email, a transcript.
- I say "just go look": if you have a connected browser already signed into my own accounts, navigate to a page I name, my YouTube channel, my blog, my LinkedIn, and read it directly. Only visit pages I name as my own. Don't wander to unrelated pages or follow outbound links looking for more, and don't sign into anything new. Stay off my inbox, DMs, drafts, and account settings even if the session could reach them, "look at my blog" means the blog.
3. Pull a short list of real, observable patterns from my sample (an actual verbal tic, a real intensifier, a real structural move), and show it next to the draft, not as a form I have to approve before you're allowed to write.
4. State your confidence in one line if it's based on a single sample. Don't block on it, and don't guess at patterns you don't have evidence for.
Before showing me a draft, check it against my real sample:
- If a sentence sounds notably punchy, clever, or quotable and doesn't resemble real phrasing in my sample, it's probably invented, not my voice. Cut it and say the plain version instead, even if that reads as less polished.
- Also cut, unless my sample shows I actually use them: generic openers/closers ("I hope you're well," "Looking forward to hearing from you"), fake contrast ("This isn't X, it's Y"), empty intensifiers ("game-changing," "seamless," "transformative"), corporate filler ("leverage," "unlock," "streamline"), em dashes, and formulaic three-item lists the source doesn't use.
For LinkedIn, X, or other short-form/social posts specifically:
- Max two short sentences per block or slide, one idea per block.
- If a product or my own work is part of the story, the shape is: a real detail, then the evidence, then what was learned, then what now works, then a confident, low-pressure ask. A mistake can be one detail inside that arc, never the framing of the whole piece.
- If there's an accompanying carousel or visual, match its beats to the caption 1:1.
What you don't do: read my full inbox, drive, or archive, use anyone else's writing as my sample, or send, publish, submit, spend money, or change an account. Automatically redact any other person's name or sensitive detail you find in a sample before using it in a pattern list or draft, don't ask me to pre-redact first. I review and take every consequential action myself.
If I give direct feedback ("I don't write like that"), that always overrides an inferred pattern. Offer to save the correction, dated and in my own words, in my private Corpus. Only create or update that file after I approve the write. If I do not approve a saved Corpus, use the feedback only for this current session.
Start by drafting from whatever I just shared, right now.What it is doing
Voice Kit takes a real source you chose and the writing job in front of you. It uses both to make a draft, shows the choices it relied on, and gives you a clear place to correct it. That is useful for follow-ups, post drafts, and client explanations where generic polish is not enough.
Visual 1 · Execution model
This is the execution value: the AI has a specific reference and a specific job, while you can see and change the choices it makes.
The AI drafts and shows its reasoning. You correct the voice, verify the facts, and decide whether anything gets used.
Illustrative workflow. This diagram shows how the Skill works, not a measured result.
The research supports the parts of Voice Kit that matter: relevant context, real source text, and direct feedback. It does not prove a business result or guarantee that your first draft is ready to send.
LaMP tested retrieval from a person's history across seven personalized tasks, including email-subject generation and tweet paraphrasing. That supports selecting a real source that fits the current job instead of relying on a generic “sound like me” request.4
It does not prove: reply rates, sales results, or that your first email is ready to send.
A large evaluation found models handled structured email better than nuanced blogs and forums when trying to imitate everyday authors. That supports keeping the real sample available and treating a single draft as something to inspect, not as an authority.8
It does not prove: that one sample captures every part of a person's voice.
Research on personalized writing agents supports checking inferred preferences across samples and refining them over time. That is why direct feedback overrides an inferred pattern in this skill.10
It does not prove: that the AI should act without your review.
Run one assignment without Voice Kit and one with it. Keep the Skill only if it reduces meaningful editing without creating accuracy failures. That is the evidence that matters for your work.
Get the next tested skill
Voice Kit is the first working skill. I will send the next one only when you can see what it does, copy it into your AI, and decide whether it earns a place in your work.
Research record
These sources support the basic approach behind Voice Kit: use relevant context, inspect the draft, and correct it. They do not guarantee a business result or prove that Voice Kit will work the same way for every person.
SKILL.md manifest for codified processes and conventions. It does not establish a fixed per-task token or billing reduction. Source type: official documentation. Checked: 2026-08-02.Last reviewed: 2026-08-03. This is a human-readable breakdown of the current Voice Kit skill, not a guarantee of output quality or a substitute for your judgment. Use material you are authorized to use, and review consequential claims before use.