Last month this system produced 90,721 impressions. 85% from people who don't follow me.
If you run something, you know you should be posting, and you open the box on a Tuesday with nothing to say, this is for you. The problem was never that you can't write. It's that nothing collects what you already said.
What follows is not how mine works. It's how to build yours. Twelve prompts, in order, pasted cold into your coding agent. Every step tells you what you should have when it finishes, so you can tell a working step from a broken one.
Nothing is gated. Finish it and you have the system.
Give it a room to live in
A folder and a memory file. Every coding agent reads one file automatically at the start of a session, whether that is CLAUDE.md, AGENTS.md, or your tool's equivalent. It's the only thing guaranteed to load every time, so everything else hangs off it.
Create the memory file your agent loads automatically at the start of every session. Work out what that file is called for the tool I'm using and tell me before you write it. I'm building a personal writing system for LinkedIn. This file is the always-loaded memory: who I am, who I'm writing for, and where everything else lives. Keep it under 60 lines. It's an index, not a document. Use a table to point at files that don't exist yet, and I'll fill them in as we build. Create the empty folders it references.
a memory file under 60 lines, plus the empty folders it points at. It should have told you which filename your tool uses. If it wrote an essay, tell it to cut it in half.
Write the spine before anything else
Every later decision is downstream of this. Who you are, who's reading, and what the writing is for. Skip it and the system produces competent posts aimed at nobody.
Interview me, one question at a time, until you can write my
positioning spine. Never give me a list of questions.
Six things have to come out of this, and you do not move on from
one until it is concrete:
WHO The ONE person I'm writing for. "Founders" is not a
reader. "The owner of a 30-person accounting firm who
has already bought AI tools nobody uses" is a reader.
WHY ME What I have actually done that they haven't. Not my
values, not my mission. The thing on my record.
PROMISE Where I'm helping them get to, and who they become.
Not features. Steve Jobs didn't say "5 gigabytes of
storage", he said "1,000 songs in your pocket".
PROOF What I can point at. Imagine I've lost my voice and
can only point at things that exist in the room. I
cannot point at "trusted by industry leaders". I can
point at the logo. I cannot point at "proven results"
I can point at the number.
WHY NOW What another year on their current path costs them.
Never scarcity, never a fake deadline. Just the
arithmetic of staying where they are.
HOW My mechanism, in the shape "everyone thinks X, the
problem is Y, so the answer is Z". If I can't name it
in one line, say so plainly. That gap shows up in
every post I write until I close it.
Push back when an answer is vague. Keep asking "so what?" until
we hit something that would make a stranger feel something.
Then give me the three or four subjects I'm allowed to write
about, drawn only from those answers, and refuse anything outside
them later even if I ask.
Write it to positioning.md. End by telling me which of the six
you're still unsure about.positioning.md names one reader specifically enough that you could picture a real person you know.
Decide what kind of post you're allowed to write
This runs before every hook and it is the decision most people never make. The wrong genre kills a post that is perfect in every other way, and you will never know why.
Write genres.md. This is the first gate every idea passes,
before any hook exists.
Four kinds of post travel. Everything else is a rounding error:
A TAKE a real opinion on a NAMED thing: a company,
a tool, a country, a practice. Arguable. People
can disagree with you in the comments.
REAL LIFE something that actually happened to me, with
the specifics intact.
REAL DRAMA a named, checkable event with real stakes. Not
a story I heard. One I can link to.
BUILD IN PUBLIC the actual numbers of the thing I'm making,
including the bad ones.
And two that reliably die, which you should refuse outright:
THE ANONYMOUS TEARDOWN "A CFO once told me his AI saved 100
hours a month, but it saved him no
money." Nobody is wronged. Nobody is
real. It reads as invented because it
usually is.
THE HOW-TO TIP "5 steps to a better AI strategy."
Nobody takes instruction from someone
they have not decided to trust yet.
<example verdict="kill and convert">
before: A partner at a mid-sized firm told me their AI pilot
failed because nobody owned it.
after: Our AI pilot failed and it was my fault. I owned it
and I never once opened it after week two.
why: same lesson. One is a composite nobody believes, the
other is a confession they cannot argue with.
</example>
Then interview me for my VEINS, the four or five recurring
storylines everything I write comes out of. Not topics. Veins.
A topic runs out; a vein keeps producing because my life keeps
feeding it.
Get them from what I actually do, not what I'd like to be known
for. Then tag every idea in the harvest with its vein, and never
run the same vein twice in a row.
The reason: I will otherwise write about whatever happened most
recently, which means the same three stories on rotation forever
while I believe I am being spontaneous.genres.md exists and refuses something. Hand it an anonymous "a CFO told me" idea. It should kill it and offer you the first-person version instead.
Point it at your week
This is the part that kills the blank page. You already produce the raw material and then never open it again: notes, voice memos, meeting transcripts, mail, chats.
Write a reusable routine called harvest that reads my raw
material and pulls out things worth writing about. Save it
however your tool saves reusable instructions, and tell me how
to run it.
Sources are the places I'm about to give you. For each one, work
out whether you can actually reach it, and say so before writing
any code. If you can't reach one, tell me and skip it rather
than pretending.
ORDER MATTERS AND IT IS NOT NEGOTIABLE. My own material first,
always: notes, voice memos, meeting transcripts, my own sent
messages. Only when that is genuinely dry do you go looking
outside. A system that reaches for the web first produces posts
anyone could have written, which is the whole failure.
Each candidate carries the exact source it came from, quoted.
An idea with no source gets dropped, never invented.
Two tests before an idea makes the list:
- Can I tell this as something that happened to me, or something
I watched happen? If it can only be told as "here is what you
should do", kill it. Nobody reads advice from a stranger.
- Is there a real person in it? Not "a CFO" or "a founder". A
named one, or me. Anonymous composites are the single worst
performing thing you can write.
Track how stale the material is. If nothing new has come in for
a week, say so at the top of the output. A dry well doesn't
announce itself, it just quietly recycles.
This is the shape I want each candidate in:
<example verdict="keep">
idea: The client paid the invoice, then never signed the
contract. Turns out those were unrelated events.
source: voice memo, 14 Aug, "she's paid the deposit but
DocuSeal still says unsigned, three weeks now"
angle: something that happened to me
person: me, and a named client
</example>
<example verdict="drop">
idea: Businesses should have clearer contract processes.
source: none
why: no source, nobody it happened to, and it can only be
told as advice from a stranger. All three are fatal
on their own.
</example>
Ask me for the sources now, one at a time.run it. You get ideas with real quotes from your own week attached. Uncited ideas are the failure mode. Kill them.
Let it learn you once
One long interview, so it never writes in the flat consultant voice. The reason it stops sounding like a machine is that it stops guessing what you think.
Interview me across the areas of my life that show up in how I
think: the work, the money, the body, the people, the beliefs.
One question per message. Never a list.
Go three whys deep on every answer before you move on. The first
answer is the rehearsed one. The second is closer. The third is
usually the one I haven't said out loud, and that is where the
posts are. When an answer sounds like something I'd say at a
networking event, tell me it sounds rehearsed and ask again.
Then look for COLLISIONS, the same truth showing up in two
different parts of my life. How I handle money and how I handle
my body. What I believe about work and what I believe about
people. Those pairs are the posts nobody else can write, because
nobody else has that exact pair. Most people mine one area at a
time and end up with the same five stories forever.
Write voice.md with:
- the words I actually use, and the ones I refuse to use
- the rhythm of how I talk. Short or long, dry or warm,
where I swear and where I go quiet
- what I have genuinely lived through, in specifics, with the
dates and the numbers where they exist
- opinions I hold that most people in my field don't
- the collisions, written as pairs
Do not summarise me into adjectives. "Direct and analytical" is
worth nothing to you later. Quote me, verbatim, at length.
And never push past something I've said I don't want to talk
about. Some of it stays mine.voice.md quotes you in your own words and contains at least one opinion you'd defend in a room that disagreed.
Choose five sources, not a feed
A trending feed makes you write what everyone else is writing, one day later. Five places you actually trust, checked while you sleep, makes you early.
Add a sources step to the harvest routine. I'll give you five sources I trust. Check them for anything new, and for each one keep the link and a real quote. Never a summary without the quote, I need the actual words. Rank what comes back by ARGUMENT DENSITY: replies divided by reactions. above 1.0 people are fighting. Steal the argument. below 0.3 people are nodding. Worthless to me. A thread with 1,600 reactions and 300 replies is agreement, and agreement has nothing in it for me to say. A thread with 64 reactions and 132 replies is a room split down the middle, and that is a post. Low reaction counts in a small room is a FEATURE, not a bug: a validated idea that forty other accounts haven't posted yet. A 20,000-reaction thread is already cooked and I'd be last there. Also flag anything where the room is genuinely divided rather than piling on one side. Two-sided is what I want. A pile-on gives me nowhere to stand. Worked example, so the maths is unambiguous: <example verdict="steal this"> "If your job can be done by an AI you're lying about your job" 64 reactions · 132 replies · density 2.06 A small room, tearing itself apart. Nobody has posted this yet. </example> <example verdict="ignore"> "New model beats the old model on benchmarks" 1,601 reactions · 296 replies · density 0.18 A big room nodding along. Forty accounts have already covered it and I would be the last one there. </example> Then tell me honestly which of my five is producing nothing, so I can replace it. Don't be diplomatic about it.
it returns links and quotes, ranked, and names the weakest source without being asked.
Build the hook engine and the shape dealer
The first line is most of the post. Roughly two in three readers never tap "see more", so the visible slice isn't a preview of your post. For most people it is your post. This step also builds the thing that stops every post you write looking like the last one.
Build hooks.md. This is the single highest-leverage file in the
system, because most readers never tap "see more". The visible
slice is not a preview of the post. For most of the audience it
IS the post.
Do not ask me to invent hook shapes and do not invent your own.
Seed the library with these sixteen, verbatim in shape. They are
reverse-engineered from roughly 900 posts across ten creators.
Fill the slots with my content, never rewrite the machine.
H1 DIALOGUE COLD OPEN
A raw spoken line, zero setup. Line two is who said it.
"Cut him. He's completely unprofessional."
Because he did not blur his Zoom background.
H2 NAMED FIGURE REVERSAL
A real named person, a big number, then the turn.
Graham Walker sold his company for $1.7 billion.
His one condition: $240 million went to the staff.
H3 NEGATION REVERSAL
Line one takes away the expected answer.
I did not study in the US.
I chose China.
H4 BARE PERSONAL STAT
One autobiographical number, no framing.
I have not had an office in three years.
My team is in Armenia, Dubai and Hong Kong.
H5 SUPERLATIVE COLON
Ultra short, ends on a colon, line two pays it off.
The biggest lie of the year:
"AI will do it for you."
H6 IMPOSSIBLE NUMBER
A hard number that sounds like it cannot be true.
Five agents replaced my entire outbound team.
Research. Writing. Follow-up. All of it.
H7 CATEGORY OBITUARY
Declare the death of the expensive thing they pay for.
R.I.P. the agency retainer.
What killed it was not the technology.
H8 INDICTMENT PLUS COST
Indict the reader's whole category with a real figure.
Most firms are paying for software nobody opens.
Here is what that line costs you a year.
H9 CONTRARIAN HOUSE RULE
A counterintuitive rule you hold. Line two is a razor.
We do not let clients send cold messages.
Ever.
H10 STATUS GAP
A big achievement undercut by a basic failure.
One founder had already built a $15M business.
He still could not get a single sales call.
H11 CURIOSITY GAP
A subject pointing at something unshown.
This is the email that ended a six figure contract.
H12 ACHIEVEMENT PLUS PARENTHETICAL
A win, humanised by an aside.
I (quietly) let an AI run our books for a month.
Here is the one thing it caught that I did not.
H13 IRONIC FLEX UNDERCUT
A proud line, demolished in line two.
I love running an agency.
Just did six figures. (in unpaid invoices)
H14 THEN NOW DIPTYCH
Two dated lines, two states, the gap is the story.
2016: could not speak in public. Had a stutter.
2026: paid to speak at sold out events.
H15 FEIGNED INCREDULITY
Perform disbelief at something true, then enumerate.
Let me get this straight.
You pay for three tools, use one, and cannot say
what any of them do.
H16 TROUBLE PLUS ESCALATOR
A named thing in trouble, then a parenthetical that
makes it worse.
That rollout your board signed off is already failing.
(And the vendor knows before you do.)
Then add the three laws every filled hook has to pass:
SHAPE. Line one runs 5 to 10 words and 33 to 55 characters, far
shorter than feels right. It opens a gap and does not close it.
Hard line break. Line two drops the twist, cost or reversal. 96
to 100% of winning posts have a line two. Specificity means a
name, a number, or a quote. Never an adjective, never a category.
SUBJECT. The subject of line one is a person, or a thing being
handed a human fate such as dying, in trouble, or humiliated. A
company doing ordinary business fails. A plural mass fails,
because nobody can climb inside a statistic. An abstraction
fails. The rescue is always the same move: whatever the company
did already happened to someone, so zoom to one of them.
STAKE. Before shaping anything, answer in ONE sentence: who is
wronged, shocked, vindicated or trapped, and what exactly does
the reader feel? If you cannot name both the person and the
emotion, it is an observation wearing a story's clothes. Refuse
to write it and tell me to find the human inside the idea.
<example verdict="works">
Deloitte just got sued for $172 million
for an AI project that never worked.
Half the Big Four are next.
pattern: H2. A named giant, a dated event, a number I can
check. Something happened TO someone.
</example>
<example verdict="dies">
Many large consultancies are facing
scrutiny over AI project delivery.
why: same fact, no name, no number, nobody it happened to.
A statement about a category. Nothing to feel, nothing to
argue with, nothing to check.
</example>
<example verdict="works">
She got fired from her agency.
For posting on LinkedIn.
pattern: H1 into a status gap. One person, one injustice,
seven words. You feel it before you finish understanding it.
</example>
<example verdict="dies">
Companies are increasingly restricting
employee social media activity.
why: true, and completely unfelt. You cannot be wronged on
behalf of a plural noun.
</example>
Now build the SHAPE DEALER, and this is the part almost nobody
does. A post has four shape decisions, not one, and you deal all
four BEFORE line one exists.
1. HOOK PATTERN which of the sixteen above
2. STRUCTURE the argument shape. Seed these nine:
story scene chronological, with one line of speech
paired proof claim, evidence, claim, evidence
x versus y two labelled paths compared
inventory spine 1/ 2/ 3/ blocks, each a capability
negation not this, not this, but THIS
before and after the same thing at two points in time
teardown here is the thing, here is why it broke
justification because, therefore, which means
myth bust everyone believes X. X is wrong. Here
is what is actually true.
3. BLOCK the ONE dominant visual run in the body.
Seed these six:
numbered run 1. 2. 3. a sequence or ranking
arrow flow cause and effect, before and after
dash stack a scannable dashed list
contrast columns two sides mirrored down the page
colon reveal three or more short labels with a
payload under each
quoted scene two or more standalone quoted lines
4. ACCENT exactly one borrowed device on top. A single
quote, or a single colon label. NEVER a
second list. Two list types in one post read
as a template.
Rotation rules, and enforce them by keeping a record:
never the same block twice running
never the same hook pattern twice in one week
never the same structure twice running
The reason this matters more than it sounds: choose well every
single time without rotating, and every post you write looks
identical. The reader stops seeing posts and starts seeing a
template, and a template is the fastest way to become
scrollable.
From now on, never invent a shape. Deal all four, name what you
dealt and what it is rotating away from, then show me the
verdict on all three laws before you write anything else.hooks.md holds named patterns with real examples. Ask for a hook and it names the pattern it used. Ask for two posts in a row and it deals a different block for the second.
Teach it to write the body, not just the hook
A great hook on a default-LLM body is still a dead post. This is the step almost every system skips, and it is why so much AI writing reads correct and lands flat.
Write craft.md. This covers everything below the fold, which is
the half nobody writes down and the half that decides whether a
good hook was worth anything.
LENGTH. Work out my real median from posts I have actually
published, and write to it. Not to what some creator with a
different audience does. If I do not have enough posts yet,
start at 900 characters and correct it once I do. Long is not
depth. It is usually two posts stuck together.
ONE IDEA. One post, one idea. If a draft carries two, tell me,
and give me the second one back as a separate draft rather than
burying it.
RHYTHM. One beat per line. A blank line under most of them.
Short lines carry weight, so put the load-bearing sentence on
its own. Vary the length or it drones. Read it aloud and fix
wherever you stumble. Squint at the shape: varied is good, one
solid rectangle is bad.
DRAFT ROUGH FROM LINE ONE. This is generative, not a polish
step. Do not write clean and then rough it up afterwards, which
produces decorated-clean and reads worse than either. Compose
with the texture already on:
drop the full stop on roughly half the lines
fragments are fine, and better than a padded sentence
one word in caps carrying the weight, not five
an aside in brackets where a real person would put one
the occasional trailing thought
Lowercase is the rarest move. Once, at a genuinely raw beat, or
not at all. Never as a default drift, which reads try-hard.
CLOSE ON ATTITUDE, NEVER A QUESTION. "What do you think?" and
"Agree?" are the weakest endings available and the platform
downranks the engagement they beg for. End on a flat statement
with a point of view, or on the consequence. Measured across my
own posts the gap is roughly seven to one.
THE $10 TEST. If every word cost me ten dollars, what would you
cut and still keep it working? Cut those words before showing
me anything.
BUT AND THEREFORE, NEVER AND THEN. Every "and then" is a place
the reader can leave.
weak: I went to the shop and then I bought milk and then
I paid and then I went home.
strong: I went to the shop, BUT I forgot my wallet,
THEREFORE I drove back, BUT someone I did not
recognise was standing in my kitchen.
CONCRETE, NOT JUST CLEAR. Clear means they understand it.
Concrete means they can picture it. You can be perfectly clear
and still forgettable.
clear but vague: we help businesses grow
concrete: took a roofing company from 12 leads a
month to 47 in sixty days
BANNED OUTRIGHT, no exceptions:
em dashes and double hyphens as separators. Use a full stop.
semicolons
emojis
exclamation marks
"it is not X, it is Y" and "not A. not B. but C."
"in today's fast-paced world", "it is worth noting that",
"when it comes to", "at the end of the day"
delve, tapestry, leverage, robust, seamless, unlock,
supercharge, game-changer, landscape as a metaphor
"experts say", "studies show", "research suggests" with no
named source
three items in a list when two make the point
rotating synonyms to sound varied. Pick one word, reuse it.
THE NAME TEST, run it on every draft: cover my name and photo.
Would anyone know this was mine? If no, it is not specific
enough yet and you have written something anyone could have
posted.hand it a draft that ends on "what do you think?" and one that runs 2,000 characters. It should refuse both, name the rule, and hand back the fix rather than asking you what you want.
Build the gates that stop bad posts
This is what separates a system from a prompt. Without gates it will happily hand you a clean, competent, completely forgettable post, and you'll publish it because nothing objected.
Add gates every draft passes before you show it to me. Print each
verdict with the reasoning. A gate that silently passes is not a
gate, it's decoration.
1. STAKE. Name the wronged party and the exact emotion in one
sentence, or kill the post. This runs FIRST and it is the
highest-leverage gate you have. A flat observation with a
perfect hook still dies on the feed.
2. SUBJECT. Who is the human in line one and what happens TO
them? An organisation doing business fails. So does a plural
mass and an abstraction.
3. SHAPE. Line one 5-10 words, opens without resolving, hard
break, line two drops.
4. BLOCK. The body needs ONE dominant visual block, a numbered
run, two mirrored sides, a dashed stack, a quoted scene. Not
one shape start to finish. Measured across ~890 posts from ten
creators: the median post uses three distinct devices, 85% use
at least two, and pure single-device posts sit in the bottom
15%. A lone quote or a single label is an ACCENT, not a block.
And never two list types in one post, pick one.
5. TEXTURE. Clean is the tell. One beat per line, with 55 to 85%
of lines at eight words or fewer and a median around 6 to 9.
Drop the full stop on roughly half the lines. One word in caps
carrying the weight. If every line is sentence-cased with
every period in place and no fragments, a machine wrote it and
the reader can feel it in two lines.
6. FLOW. Hunt every "and then". Replace with "but" and
"therefore". "I went to the shop and then I bought milk and
then I went home" has nothing pulling you forward. "I went to
the shop, BUT I forgot my wallet, THEREFORE I drove back, BUT
someone I didn't recognise was standing in my kitchen" does.
Every "and then" is a place the reader can leave.
7. COST. If every word cost me ten dollars, what would you cut
and still keep it working? Cut those words.
8. TRUTH. Every number true and sourced. No invented figures,
no invented people, no rounded-up claims. One made-up number
is unrecoverable, it's the only mistake that costs the whole
account.
Show me all eight verdicts under the draft, every time.
Here is what a real verdict block looks like, so you know the
level of specificity I want. Vague verdicts are worse than none,
because they read as passing.
<example type="a good verdict block">
1 STAKE PASS, a bookkeeper is quietly blamed for an
error the software made. Reader feels the unfair
bit, because they have been that person.
2 SUBJECT PASS, "She reconciled it three times", one
human, present in line one.
3 SHAPE FAIL, line one is 14 words and resolves itself.
Cut to "She reconciled it three times." and move
the reason to line two.
4 BLOCK PASS, numbered run of four, dominant, no second
list competing with it.
5 TEXTURE FAIL, 22 of 24 lines end in a full stop and every
line is sentence-cased. Reads machine-written.
Drop the stop on roughly half.
6 FLOW FAIL, three "and then"s in the middle section.
Convert to but / therefore.
7 COST "in order to" x2, "it is important to note" x1.
Cut, saves 14 words.
8 TRUTH PASS, the one figure is sourced to the invoice.
</example>
<example type="a useless verdict block">
Stake: good. Subject: fine. Shape: looks right.
Texture: reads well. Truth: no issues.
why it's useless: no evidence, no line quoted, nothing I can
act on. A gate that cannot fail is decoration.
</example>
Fail loudly and specifically. I would rather rewrite three times
than publish something that quietly passed.give it a deliberately flat idea. It should refuse it at gate 1 rather than write it beautifully.
Program the week, don't wait for inspiration
Everything so far produces one good post when you feel like it. This is what turns it into four a week whether you feel like it or not, which is the only variable that actually compounds.
Build the week programmer, and a tracker underneath it.
THE PROGRAMMER. Given my target number of posts a week, plan
them from the harvest before I write any of them. For each slot:
which vein, which genre, which hook pattern, which block. Deal
the whole week at once so the rotation is visible and you cannot
accidentally run the same vein twice.
Weight it. Most slots go to the posts that reach people who
don't know me. One slot per week goes to the post aimed at
someone who might actually buy. Never invert that ratio, and
never let a week be all one or all the other.
THE TRACKER. Read the state of my week off the actual files, not
off memory, and open every session with it in five lines or
fewer. It has to be able to tell me things I do not want to hear:
behind posted 1 of 4, it's Saturday, here's the gap
repeating the last two posts ran the same vein, rotate
stale nothing new in the material for 9 days
unmeasured 11 posts carry no numbers, so no median exists
and I cannot tell you what is working
no ask 8 posts and not one of them asked for anything
<example type="a good open">
Week 34 · 1 posted, 1 drafted, floor is 4
Behind: need 2 more, and it's Friday
Rotation: last two were both the same vein, pick another
Material: 9 days stale, refill before drafting
Next: draft the Tuesday slot from the client-work vein
</example>
<example type="a useless open">
You're making good progress this week. Let me know
what you'd like to work on.
why: tells me nothing I did not already know, and asks me
to do the thinking the system exists to do.
</example>
Two rules that make this survive contact with a bad week:
Refuse to draft on stale material. If nothing new has come in,
say so and make me refill first. A system fed nothing recycles
what it already said and calls it output.
Never scold. State the gap in one line and move on. If it
lectures me I will stop opening it, and a system I do not open
is worth nothing.run it on a week where you've posted nothing. It should say so plainly in the first line, name the gap, and refuse to draft if your material has gone stale.
Put it behind one trigger
Everything above is parts. This is the part where you stop thinking about it. One trigger, and the finished post comes back.
Make the whole line runnable in one go, in order: read my week, check my sources, pick the idea, choose a hook pattern, draft it in my voice, run every gate, then show me the draft with the verdicts underneath it. Save it the same way you saved harvest, and tell me exactly what to type to run it. Open by telling me where I stand this week in five lines or fewer, before you do any work: what's posted, what's drafted, how stale my material is, and the one thing to do next. Deal the shape first, hook pattern and structure and block and accent, and tell me what you dealt and what it's rotating away from. Then write. Never the other way round: a body written first and a shape checked afterwards always comes out as one flat ladder. If the raw material is stale, say so and refuse to draft until I refill it. Recycling old material quietly is the failure I want you to make impossible. Never publish. Show me. I post it myself.
one trigger produces a status read, a draft, and visible verdicts. If it publishes anything by itself, take that ability away.
Feed the numbers back in
Without this you have a machine optimising for nothing. Most people stop here, keep posting, and never learn which posts worked.
Every post gets two numbers written back into its own file:
impressions, and conversations. A conversation is a DM, a reply
thread, or a call that came out of that post. A real person
talking to me about actual work.
Hold these apart, because mixing them is how people quit good
writing. Impressions are the PROXY. Conversations are the THING.
A post with 1,200 impressions that started two real conversations
beat a post with 40,000 that started none.
Grain matters too. Judge craft per post. Judge the goal every
eight posts. One post starting nothing is noise, eight posts
starting nothing is signal.
When eight posts produce zero conversations, do NOT tell me to
write better hooks. If the gates are passing, craft is not the
constraint and rewriting hooks is the comfortable work. Tell me
to change the room I'm writing to, or change what I'm asking for.
Once I have enough posts, work out my median and name which ones
genuinely beat it. Then tell me what those had in common, and be
specific. "Stronger hooks" is not an answer. "All four opened on
a named person in trouble" is.
Never estimate a number to fill a gap. Blank beats wrong. A
guessed number makes a dead funnel look alive and quietly poisons
every conclusion after it.
The level of specificity I want when you report back:
<example type="a useful finding">
Four posts beat your median of 1,400.
All four opened on one named person in trouble.
The five below median all opened on a category
("most firms", "a lot of founders", "companies").
Your two highest also carried a checkable number in line two.
</example>
<example type="a useless finding">
Your stronger posts had better hooks and more
engaging openings. Consider leaning into that.
why: I cannot do anything differently on Monday morning
because of that sentence.
</example>after eight or so posts it can tell you which beat your median and why. If it can't, you haven't been logging.
What this won't give you
- It won't find your positioning for you. Step 2 is an interview, not an oracle. If you don't
- It won't survive being ignored. The harvest goes stale in about a week. A system fed nothing
- It won't make you consistent. It removes the excuse, not the decision.
- It won't do this for your team. One person's voice, one person's week. Installing this
Don't want to build it yourself?
One working session and it is running. We set the whole thing up inside your own coding agent, on your machine, in your voice. You own it when we leave, and there is nothing to subscribe to.
Book a call