Founder Voice Email Writing

Related reviewer (if you have one)

If you've configured a separate agent to review outreach emails for voice quality, point it at this skill as its source of truth for degradation patterns — have it read this file at the start of every review, and cross-reference back to it by name when it flags a clause that violates a pattern documented here. This is optional; the checks below work fine done by the same agent that writes the email.

Before you start: whose voice, and do you have samples of it?

This skill exists to protect one specific person's real writing voice from being smoothed into generic AI prose. That person is whoever you're writing on behalf of — check first whether you have real samples of their own words to work from:

1. Look for alignment-harness records voice (run that command to get the folder) and a samples.md file inside it.
2. If it exists, that's your canonical source — read it before writing anything.
3. If it doesn't exist yet, say so plainly and offer two things: (a) ask them to paste or dictate 3-5 sentences of their own writing — enough to start calibrating, the same way a first draft of any verbatim source gets built — or (b) if they'd rather see the pattern in action before supplying their own, offer the worked example from examples/author-approved-samples.md in this skill folder, clearly labelled as a borrowed starting point to replace, not a template for their own voice. (That file ships only once a small set of approved writing is published for that purpose — none yet. If it is missing, say so and go with option (a).)

Everything below is the method. Apply it to whichever voice you're actually protecting.

The lesson this method is built on

An AI agent once rewrote a person's outreach emails for a client-facing campaign. Their response: "almost every edit you made was worse." Every time the AI "improved" the words — tightening sentences, adding transitions, making it "flow better" — it stripped authenticity and added sales language. The emails went from sounding like a real person to sounding like an AI trying to sound like a real person. (A fuller, worked version of this story ships as an opt-in example once it's approved for publication; none is published yet.)

The principle

The person's verbatim words with grammar fixes only are ALWAYS better than any AI rewrite.

This is not a style preference. It's a hard constraint. The person's voice is the outreach's voice. When the voice sounds artificial, the whole effort fails — because it depends on feeling like a real person reaching out, not a campaign.

Where this skill operates from

The principle above — verbatim words with grammar fixes only — is the operational stance. The reason it works is what's underneath it: an email carries someone's voice because it comes from their actual intent in writing it, not just from its surface wording. So the preservation work isn't only about avoiding rewrites; it's about recognizing that the substrate of the email is a person having something they want to say to someone specific, articulated in their particular way for that moment.

When agent-side work touches that artifact, it's the substrate that gets preserved or distorted, not just the surface. Grammar fixes preserve the substrate because they correct accidental friction without changing the intent it carries. "Improvements" — tightening sentences, adding transitions, making the prose flow better — distort the substrate because they replace something that came from the person with something that came from an agent's pattern of what person-shaped writing looks like. The recipient detects the substrate change, not the wording change.

The same consideration applies to sourcing material for new email work. Canonical material — someone's own dictations, things they've authored, their raw transcript voice samples — carries their substrate. Material derived from that source (a review report about their voice, an agent's character sketch of them, agent-paraphrased passages, prior agent prose they once approved) carries their patterns but not their substrate, because the synthesis step separates the words from the intent that originally produced them. The first kind of material can be selected from and arranged. The second kind can be referenced for facts but not consumed as "how they write."

When the email work needs material that doesn't exist yet in canonical form, the move is to ask for a sample — three sentences is enough to start — rather than compose toward what you imagine their dictation would have produced.

Related skills: writing-aligned-user-facing-words (the equivalent stance for public-facing website copy, governed by the same substrate principle); how-to-talk-like-the-founder (if you have it — the fuller version of this same voice-preservation discipline, covering more than email).

What "grammar fixes only" means

• Fix obvious typos and spelling errors
• Fix subject-verb agreement if clearly wrong
• Fix sentence fragments only if they're genuinely confusing (most are intentional rhythm)
• Capitalize proper nouns
• That's it. Stop there.

What you must NOT do

• Do not "tighten" sentences
• Do not add transitions ("Additionally," "Furthermore," "That said,")
• Do not restructure paragraphs for "flow"
• Do not replace their word choices with "better" ones
• Do not add qualifiers or hedges they didn't use
• Do not remove repetition (repetition is often intentional emphasis)
• Do not add bullet points or formatting to prose they wrote as paragraphs
• Do not add a "professional" sign-off if they just wrote their name
• Do not clean up their punctuation style (em dashes, ellipses, sentence fragments)

How to use this skill

When the person gives you spoken or written words to turn into an email:

1. Transcribe or copy their words exactly
2. Fix only grammar (typos, agreement, capitalization)
3. Present the result
4. Do NOT offer "polished" or "tightened" alternatives
5. If they ask you to change something specific, change only that thing

When writing new outreach from scratch (no sample from them yet):

1. Ask the person to speak or write the core message first — this is always better than the agent guessing at their voice
2. If that's genuinely not possible right now, write in the plainest possible language — short sentences, first person, conversational — and say clearly that this is a placeholder draft, not their calibrated voice
3. Read it aloud. If it sounds like a LinkedIn post or a marketing email, rewrite it simpler
4. Avoid: "leverage," "optimize," "solution," "streamline," "innovative," "cutting-edge," "transform," "empower," "unlock potential," "drive results"
5. Prefer: "help," "fix," "find," "build," "figure out," "what's blocking you," "what's working"

When reviewing existing copy for voice:

Compare against a real sample of the person's own writing (or, absent one, the approved example in examples/author-approved-samples.md if it exists — none yet). Ask:

• Does this sound like a person talking, or a campaign?
• Would you send this to a friend, or does it feel like it came from a CRM?
• Are there phrases no human would say in conversation?
• Is there any language that positions the sender as above the reader rather than alongside them?

The progression this method comes from

This is the shape the lesson usually takes, illustrated by a real worked example (the full version ships only once it's approved for publication):

1. AI writes marketing-style copy with sales language, booking links, product pitches
2. The person says: that's not what this is
3. AI rewrites with better framing but still in AI voice — polished, structured, corporate-casual
4. The person says: use MY words, just fix the grammar
5. The grammar-fixed version of their actual spoken words turns out to be the best one in the sequence
6. Lesson locked: the person's own voice beats AI voice, always, no exceptions