50,000 tweets. One file. Every pattern extracted.
6 years of writing. 50,796 tweets scanned. Filtered to the 13,962 that performed. Reverse-engineered into the rules, constraints, and word-level mechanics that made them work.
Not a style guide. A system file. Drop it into any AI and generate content in this voice.
SOURCE
50,796 tweets scanned
13,962 indexed (RTs, replies, URL-only removed)
SHAPE
median length 9 words
under 10 words 57%
single sentence 78%
shorter = more likes (258 avg at 6-10w, 139 at 31+w)
PERSPECTIVE
"you" vs "I" 30% vs 7%
"we" 0%
PUNCTUATION
line breaks/tweet 1.3
periods/tweet 0.7
question marks 0.04
exclamation points 0.01
STRUCTURE
observation/declaration 82% of openers
numbered list 6%
conditional "if" 4%
imperative verb 4%
6 RHETORICAL MOVES
contrast pairs renting/owning, theory/practice
reframes flip what you're doing into what you should be doing
math quantify abstract ideas (+2 vs x2)
uncomfortable truths say what people avoid saying
compressed wisdom entire frameworks in two lines
deadpan humor through understatement
12 CONTRAST FRAMES
reframe 23% "Overthinkers are underpaid."
parallel declaration 17% "X does A. Y does B."
paradox 12% "Customers that pay more, complain less."
conditional reveal 11% "If X, [surprising Y]."
juxtaposed pair 7% "distraction, focus" (no verb)
progression 7% numbered list, each step escalates
explicit vs 6% "renting vs. owning"
negation flip 4% "No one cares... everyone cares..."
+ 4 more rare frames (expectation subversion, labeled, chiastic, cyclical)
11 WORD-LEVEL MECHANICS
alliterative contrast same-letter opposites (complexity/clarity, default/design)
matched meter couplets with equal syllable counts (9/9, 4/4)
chiasmus A-B flips to B-A
circular loops ending returns to beginning
internal rhyme learn/earn, build/billed
negation flips same words + "don't" inverts meaning
paradox contradicts itself to reveal truth
monosyllabic endings punchlines land on one-syllable words
lowercase as register "thinking out loud" > "making a pronouncement"
drop the period bare-word endings get ~20% more engagement
land on a noun final word is a thing, not an action
VERB MOOD
declarative 80% asserts. the default.
imperative 9% commands. "Build distribution."
conditional 9% "if" setups. slightly outperforms imperatives.
interrogative 2% rare but highest ceiling.
CLOSING PATTERNS
lands on a noun, no period, declarative statement
punchline inversion ~10% (short line after long setup)
COLON AS PIVOT
17% of top tweets use a colon
+33% retweets vs non-colon tweets
both sides 2-4 words
10 CATEGORIES OF SILENCE
no self, no diary, no ask, no performance, no tribe
no complaint, no cliches, no own metrics, no pop culture, no news
40+ BANNED WORDS
no synergy, no ecosystem, no stakeholders, no thought leadership
10 REWRITE PAIRS
same idea, generic vs. Jack - teaches the compression by example
29 REFERENCE TWEETS
highest performers, sorted by engagement
Paste jackbutcher.md into any AI conversation. Prompt as usual.
[paste jackbutcher.md]
Write a post about compounding.
The constraints do the work.
The file isn't the moat. The person is.