---
name: "executive-cookbook-voice-kit"
version: "2.1.0"
description: "AI Executive CookBook: draft writing that sounds like the owner, not generic AI, from a real sample they share, a named public page, or their own past conversations. On first activation in an environment, defaults to reading past conversations as the fastest starting point, then offers to draft. Use when someone wants an email, post, or explanation to start from their actual words, with visible pattern notes they can correct. Mechanism only, no one person's data lives in this file, that stays in the owner's own profile file."
---

# Voice Kit

Part of the [AI Executive CookBook](https://why57.com/ai-execution-cookbook.html). Shared logic only, no one person's samples or corrections live here, those go in the owner's own profile file (their Corpus).

Make the model write like a specific person, not generic AI. Give it real samples of that person's actual writing or speech, and check anything punchy or quotable against those real words before using it.

**The one idea that matters:** abstract rules ("be direct," "avoid jargon") produce AI-sounding output, because the model invents phrasing that satisfies the rule instead of recalling the owner's actual words. The fix: keep real sample text on hand, and check anything that sounds notably punchy against it before using it.

## Starting a new Voice Kit

Run this once, the first time this skill activates in a given environment, before drafting anything, even if the owner hasn't asked for a draft yet.

If this environment can list and read the owner's own past conversations, that's the default, fastest starting point, offer it directly instead of an open-ended question:

> Fastest way to get your voice: I can read back through what you've already typed to me here and pull real patterns from that. Want me to start there, or do you have a sample or page you'd rather use instead?

- **Default: past conversations.** If the owner agrees, or says nothing more specific, pull their own messages from past sessions as a starting source, not training data, and draft from them.
- **A sample they give you instead.** A link, a doc, a pasted email, a transcript, a paragraph. One sample is enough, don't ask for more before drafting.
- **A named public page instead.** If a connected browser is already signed into the owner's own accounts and they name a page as theirs (their blog, their YouTube channel, their LinkedIn), go read it directly instead of asking them to paste it.
- **None of this is available.** Ask for one sample directly and draft from that.

Once a starting source exists, draft from it immediately, don't run a full intake interview. After the first draft, offer Calibration Mode as the next step: "Want to add another sample or correct anything, to sharpen this further?"

## Show your work

Pull a few real, observable patterns from the source (a verbal tic, a real intensifier, a structural move) and show them beside the draft so the owner can correct anything off. State your confidence in one line if it rests on a single sample.

## Check the draft before showing it

Scan for anything notably punchy, clever, or quotable. If it doesn't resemble real phrasing in the source, it's probably invented, cut it and use the plain version instead. Also cut, unless the source shows the owner actually uses them: generic openers/closers, fake contrast ("This isn't X, it's Y"), empty intensifiers, corporate filler, em dashes, formulaic three-item lists.

## Short-form needs its own shape

For LinkedIn, X, or similar: max two short sentences per block, one idea per block. If the owner's own work is part of the story, the shape is a real detail, the evidence, what was learned, what now works, then a confident low-pressure close, never a scripted call-to-action. A mistake can be one detail inside that arc, never the frame of the whole piece. Match an accompanying carousel to the caption's beats 1:1.

## Keep the owner's data in their own Corpus file, never in this skill

This skill is the logic. The owner's real samples and their dated corrections log belong in a separate file they control, for example `<owner>-voice-corpus.md` in their own project folder.

- If none exists, create one and say so.
- Before a short-form or high-stakes draft, read it directly rather than working from memory.
- Log every correction there, dated, in the owner's own words.
- Never paste this file's contents into this skill, into an outward-facing draft, or into anyone else's Corpus file.

## Calibration Mode (say "let's calibrate")

A repeatable way to grow the owner's Corpus file on demand, instead of only improving by accident. This is also the natural next step right after onboarding, once there's a first draft to react to.

1. Ask for a new sample, spoken or written, whichever is thinner. If nothing new, work from the existing Corpus.
2. Draft a quick test targeting whatever's weakest.
3. Ask directly: "Does this sound like you? If not, what's off?" Push for something specific, not a vague reaction.
4. Log the answer to the Corpus file immediately, dated, verbatim.
5. Offer another round right away.

Direct owner feedback always overrides an inferred pattern.

## What this skill doesn't do

Doesn't send, publish, submit, spend money, or change an account, the owner reviews and takes every consequential action themselves. Doesn't claim an exact voice clone or invent facts. Beyond that, what's appropriate to read, store, or repeat is a judgment call for the harness you're running in, not a rule this skill needs to spell out separately.