AI content workflows that don’t ruin your founder voice

You can spot AI-written founder content from a mile away. Same rhythm, same em-dashes (the wrong kind), same tidy three-bullet structure, same vague closer about "the future." The audience has learned to skim past it. Here’s how to use AI without ending up in that pile.

Why AI flattens voice by default

LLMs regress to the mean. Their job is to produce the most likely sentence given the prompt — which is, by definition, the average. Your voice is what makes you not average. If you let the model write first, you get average. The fix isn’t a better model; it’s a different workflow.

The workflow that keeps voice

• You write the spine. Three to five sentences, in your own voice, in a Google Doc or even Apple Notes. The hook, the core argument, the closer. This is the part that has to be you.
• The model fills the middle. Pass it the spine plus three of your past posts and ask it to draft the connecting paragraphs. Specifically tell it to match the rhythm of the examples, not just the topic.
• You edit out the AI tells.Cut every "in today’s landscape," every "let’s dive in," every paragraph that ends with a rhetorical question. Tighten anything that sounds suspiciously balanced.
• Read it out loud.If it doesn’t sound like you talking, it isn’t. Rewrite the bits that fail this test.

The prompt that actually works

Generic prompts give generic output. The prompt that holds voice looks like: "Here are three of my past posts. Match the sentence length, the rhythm, the way I open and close. Don’t use the words [insert your AI tells]. Draft the body of the post below based on this spine." Specificity is the whole game. Vague prompts equal beige output.

The repurposing workflow that doesn’t suck

Long form to short form is where AI earns its keep, but only if you’re not letting it write the hooks. Have it pull five passages from a long post that could stand alone, then you write the hook and the closer. The model picks; you frame. This split keeps the voice intact and saves about 80% of the time.

What to never let the model do

• Open the post.Hooks are the highest-leverage sentence you write. Nobody’s second-best hook beats your worst hook. Write it yourself.
• Make claims with numbers. Hallucinated stats are a brand risk. If a number matters, you sourced it.
• Write the close.The last line is what readers remember. A model close is generic by construction. Yours shouldn’t be.

The detection question

Your audience can tell. Maybe not consciously, but reply rate and share rate quietly drop on AI-flattened content. We’ve watched founder accounts in SF and Brooklyn lose 40% of engagement over a quarter after fully outsourcing to AI. The accounts that kept voice and used AI for plumbing held steady or grew.

The honest output split

For most founders we work with, the right split is roughly: 70% you, 30% AI — and the AI 30% is research, structure, repurposing, and editing. Not first drafts. Not opens. Not closes. If your current split is 30/70 the other way, that’s why your content stopped working.

How we help at The Nerdish Mic

We design AI content workflows for US founder-led brands that keep the founder’s voice intact. Prompt systems, editing guardrails, repurposing pipelines — the boring scaffolding that lets you ship more without sounding like everyone else. If your content has gone beige, we can fix that.