# You Valued Sending More Than Writing

By Wes Pomeroy. Published April 25, 2026. https://wespomeroy.com/writing/ai-written-documents-at-work/

I revised an overview document Claude had drafted about my own AI work. Pulling out the tics. Rewriting sentences that didn't sound like me. Not because the content was wrong — Claude got the content right. But if I ever had to speak any of this out loud, it couldn't sound like what I wrote and what I'd say came from two different people.

Then I remembered a message I'd sent to a colleague. I'd asked Claude to draft it. I hadn't really read it. I hit send.

The overview I cleaned up. The message I didn't. Both were drafts I hadn't really written.

What got me to slow down wasn't the message. It was a podcast.

I've been listening to one I'm pretty sure is AI-generated — the hosts, the script, all of it. After a couple of episodes the tells started piling up. No callbacks across episodes. Uniform pacing where a real conversation would breathe. Filenames mispronounced in ways no human would mispronounce them. Once I started counting, I had more than a dozen.

And that's when the message to my colleague came back to me. Same kind of off. Just quieter.

## Different physics

We hand off communication all the time. "Please email so-and-so and tell them X." Most of us have a feel for how much to trust the result. With a junior teammate, you know there's risk and you mitigate accordingly — ask for a CC, a draft to review, a heads-up before something important goes out. Your read of them sharpens over time. You learn what they're good at, where to keep an eye, when you can step back.

AI doesn't work that way. The mistakes it makes don't have the same form as a person's mistakes. A person's wrong notes have causes you can read — they were tired, didn't have context, assumed something. AI's misses come from places you can't see. A pattern in training data. A phrase from a different register the model thought was close enough. A context cue it took the wrong way. You can't build a feel for AI mistakes the way you can with a person's. So you can't get to "I trust this 90%, skim and send" — not because the model is unreliable, but because you don't know what its 10% looks like.

It's a different kind of organism. We're still learning how to work with one. It has its own patterns: strengths, blind spots, a particular kind of confidence we're still mapping. It doesn't have a growth curve in the human sense — no linear progression where each session builds on the last. The way you handle risk with it has to be different than with a person, because what you're handling is different.

The handoff analogy almost works and doesn't quite. With a teammate, you build trust by working with them — the relationship deepens, the read sharpens, the briefings get shorter. With AI you build trust by pre-loading context — your style, your voice, what to avoid, who the message is for, what's at stake. AI reads the five-page style guide every single time you ask it to write. A junior teammate would skim it once and pick up a rough version of it over months. The advantage is consistency, not memory. The trust you build is in the file, not in the collaborator.

## What actually works

Two things, both required.

**Pre-load context.** A file with your style, voice notes, what not to do, examples of what good sounds like in your voice. Most of us already have this in our heads. Putting it in a file the AI reads every time is the most useful single thing you can do.

**Read it before you send it.** Even with the best context loaded, the misses aren't predictable. The pre-load makes most drafts good enough to skim. It doesn't make any individual draft safe to send unread.

The first thing alone isn't enough. The second thing alone is exhausting. Both is the actual practice.

When I don't review, here's what creeps into my sent mail:

\- Over-glossing peer terms — explaining things to people who already know them

\- Closers that aren't mine ("if anything is unclear or additional documentation is required")

\- Passive voice

\- Sentences with neat parallel halves where I'd normally write something shorter and lopsided

\- Tone that's warm in a generic way, not warm in the way I'm warm

None of these are disasters. They're texture loss. The reader doesn't know exactly what's off. They just feel a little less like they're talking to me.

## The bigger trap

When you send unedited AI, you're not just risking factual misses. You're communicating, underneath the message, that you valued sending more than writing. Readers feel that. They can't always say what's off, but they register it. The note lands without weight. The relationship gets a little thinner. Do this enough times and you've trained everyone you write to to read your messages with less attention, because there's less of you in them to attend to.

The bigger version of this is worse. If a team writes everything with AI, the people on that team don't develop the writing-as-thinking muscle that the writing was a proxy for. Junior person drafts, senior reviews, junior picks up how the senior thinks, and in five years junior is the senior. AI-as-default cuts that loop. You end up with senior people who can edit AI but can't generate. That's a different kind of senior than the one you'd have had.

And there's a company-voice version of this. A company's voice is one of the few things that's actually theirs, and it doesn't come from a brand guidelines doc. It comes from how people there write to each other. The Slack messages, the recap emails, the "hey just looping you in" notes. That texture is the company. AI defaults flatten it toward a shared middle. Not because the AI is bad. Because the AI is averaged. Every company that runs everything through the same models ends up sounding like every other one that does.

The voice file matters more than I thought.

Listing what you don't want is the easy part. I have a running file of Claude tics I've caught — phrases that sound like AI, structural moves AI reaches for, voice drifts to watch for. That kind of file is easy to grow. You hear a tic, you name it, you add it.

Listing what you do want is harder. You can't write down "be specific" or "be warm" as rules — those aren't the kind of thing that translate. Voice lives in examples, not abstractions. I'm still working out what my version of this file looks like. What I know so far: feeding the AI samples of what you actually sound like gets it closer than any list of style instructions can.

## What I'm changing

For anything that goes out under my name, even a quick reply, I'm reading every line before I send it. The rule I'm trying to hold to: if a sentence could have come from anyone, that's the one to rewrite.

The pre-load and the read-before-send are the practice. Skip either and you're sending notes that aren't really from you. People can't always say what's missing, but they feel it. They start reading you with less attention. The relationship doesn't deepen. The news doesn't land with full weight. The trust gets thinner.

This is small per instance. The cost compounds. By the time you notice nobody's reading you the same way, the texture they used to find in your messages is already gone.

The only way out is to slow down at the door.