AI newsletter writing

Universal Writing Standards That All Flyletter Agents Follow

Flyletter agents follow a set of Universal Writing Standards as a quality floor when writing so they don't sound like AI. You can follow these standards, too.

Evan Tarver

Evan Tarver

6 min read

Good writing has a set of universal rules. AI breaks the same ones every time.

You know what I'm talking about. You read a paragraph and something's off before you can say what. The words are grammatical, the point is fine, but it sounds like a robot doing an impression of a person.

It's all the AI tells, of course, and then some. Sentence constructions like "here's the thing," or "it's not X, it's Y," or em dashes and words like "unleash."

Luckily, you can train AI to actually write like a human if you teach it a set of standards that good writing follows, and a set of AI tells that break those standards.

That's how I constructed Flyletter. And lucky for you, I'm here to tell you exactly how I built the Universal Writing Standards that all Flyletter agents follow so they don't sound like AI.

For a full breakdown on how to get AI to write like you, check out this guide on how Flyletter is engineered to write in your voice.

1. The Quality Writing Floor

The Universal Writing Standards starts with quality rules that all good writing follows. Rather than starting off by teaching the model how not to write like AI, it's better to first teach it some best practices on how to write like an actual human.

It doesn't matter what your unique voice sounds like, these universal truths are the minimum quality bar you should be training your AI to write on.

They break into specific groups:

Sentence Craft

Vary your sentence length on purpose. Short sentences punch. Medium ones carry the explanation. Long ones build momentum, but only when they earn it.

Prefer active voice and present tense because they read with more authority. Cut gerunds when a plain verb does the job (write "start by mapping" not "starting by mapping"), and watch the fragment trap. A fragment can land hard for emphasis, but three fragments in a row is a model faking rhythm that real sentence variation should be doing.

Concision

Every word earns its place. If deleting a word doesn't change the meaning, delete it. Cut throat-clearing openers like "it's worth noting that." Cut hedging when you're not actually uncertain. Cut redundant pairs like "each and every."

Paragraph Discipline

One idea per paragraph. When you start a second point, start a new paragraph. A standalone one-sentence paragraph is powerful, so use it when a thought deserves the space. Vary your openers so two lines in a row don't start the same way.

Formatting Discipline

Every bolded term, every header, and every list has to serve the content. A bold word should surprise the eye, not numb it. Decoration for its own sake is noise.

That's the floor. Every piece of writing, human or machine, should clear it before anyone reads it.

2. Story and Anecdote Sourcing

The next part of the Universal Writing Standards teaches AI how to treat personal stories and anecdotes so there aren't any hallucinations. A fabricated story or invented statistic destroys credibility the instant a reader recognizes it.

So the anti-fabrication rule has to be tight. Every anecdote has to trace back to one of three things: something the writer actually said or lived, a documented event or statistic you can verify, or a hypothetical the reader clearly recognizes as illustrative ("imagine a founder who..."). Anything else is invention.

You need to build trust when you publish with AI-assisted writing. Firm instructions on how to deal with personal stories and anecdotes so the AI doesn't hallucinate things that aren't true are an important standard to set. Trust is the thing you can't rebuild once it's gone.

3. Top AI Giveaways

Next, the Universal Writing Standards show the AI what common tells give away AI writing and how to avoid them. It starts with the three most common AI tells:

  1. Stage-direction phrasing: The "here's the thing" move, where the writing announces it's about to make a point instead of just making it. A real writer makes the point.
  2. The em dash at machine rate: People use em dashes. Machines use them constantly. The rate itself is the tell.
  3. The colon as a dramatic pause: A colon before a standalone clause is an AI model trying to sound dramatic. Colons belong before lists and time references, not before a sentence that could stand on its own.

Then there are the pattern families, which are the same tells repeating themselves in different forms, and help the AI catch patterns instead of needing a robust list of specific words.

For example, the contrast reveal ("this isn't X, it's Y") used as a reflex. The sweeping scene-setter that opens on "in today's world." The dramatic imperative opener barking "picture this." Filler transitions like "moreover" and synonym stacking, where three near-identical words pad a line one precise word would carry.

After the pattern families are the specific phrases that flag AI on sight. The hype verbs like "unleash," for example. I won't list them all here, partly because the list keeps growing, but mostly because you know them when you see them yourself.

Which ones do you notice the most?

4. The Human Test

Even though it's important to weed out the AI tells, trying to spot evolving AI giveaways as the models evolve is like playing whack-a-mole. It's a never-ending battle of adding new words and phrases to your banned list.

As a fallback you can have the AI ask itself one question: would a busy, smart human write it this way? If the honest answer is "only an AI would write that," rewrite it. That single test catches most of the tells at once, because they all share the same root.

Good writing standards should be enforced holistically by each agent, and not fall on you to call out each and every time.

Bottom Line

Every Flyletter agent follows this set of Universal Writing Standards before writing a word. At minimum, each agent adheres to these rules so it writes like a human and not like AI.