# WRITING WELL — how to write so it doesn't read like AI

This is about how the copy READS. It is SEPARATE from the feel-seen rubric (which is
about what to recognize and say). These are the style rules. Re-read your output
against them before returning it.

1. **No em dashes (—).** The single clearest tell a machine wrote it. Use a period,
   a comma, or a colon; rewrite the sentence if you have to. Target zero; at most one
   in a whole piece, and only if nothing else works. Same for the spaced hyphen used
   as a dash. Just start a new sentence.

2. **Never invent.** Do not state an event, action, or outcome as fact unless it is
   literally in the evidence you were given. Do not infer that something happened from
   a feeling or a fragment ("he called her family", "it worked out"). When unsure it
   happened, leave it out. A true smaller specific beats an invented bigger one.

3. **No jargon, no bare proper nouns.** Plain words a smart friend with no background
   understands. If you must use a term, unpack it in the same breath. (Naming a famous
   person for stature is fine; bare domain jargon is not.)

4. **No fluff or throat-clearing.** Every sentence carries real, specific content.
   Cut "here's the thing", "the truth is", "what's remarkable is". If a clause can be
   deleted with no loss of meaning, delete it.

5. **Plain declarative sentences.** Don't stack clause on clause. Short and clear
   beats ornate. Never ship a half-formed garble; if a sentence isn't clean and
   grammatical, rewrite it.

6. **Avoid the stock AI cadences.** Don't lean on "it's not X, it's Y", "isn't just
   X, it's Y", rule-of-three rhythms, "the kind of X that Y", or a dramatic one-line
   paragraph for effect. Once is fine; as a pattern it reads as slop.

7. **Don't over-emphasize.** Bold or italic only where it genuinely helps.

8. **Reach for the concrete.** Where it fits, back a claim with a real, specific
   instance of their behavior (a move, a moment, a line they wrote) instead of a bare
   adjective. A claim with an example behind it lands harder than a floating one.
   This is a preference, not a hard gate: don't force an example where the format is a
   single tight line and it won't fit, and never pad with a weak or invented one.

9. **Prefer real names by tier.** When you name someone's caliber or where they could
   go, prefer actual, currently-desirable places matched to their field and level
   (Anthropic, OpenAI, Google DeepMind for frontier AI; Jane Street, Citadel for
   quant; Google, Nvidia for elite engineering; or a fundable founder — see
   top-companies.md) over vague placeholders like "a serious company with serious
   backing". Only name what the evidence genuinely supports.

10. **No grandiose or unfalsifiable claims.** Recognition has to be true and
    checkable. Avoid savior / once-in-history framing and sweeping absolutes ("the
    world would be safe in your hands", "incorruptible", "you will change
    everything"). Everyone is fallible; an over-claim breaks trust instantly, and it
    is not what makes someone feel seen anyway. Stay genuine and grounded in who they
    actually are. Reach high, but only as far as the evidence holds.

The test: read it aloud. If it sounds like a human who knows this person wrote it
plainly, it passes. If it sounds like a model performing insight, rewrite it.