Zkitszo - The Real
October 19, 2026

Official, Measured, Mine, Folklore

One idea, written up properly for LinkedIn. Then the X version.. the Threads version.. by the fourth one the idea is boring us.. the fifth comes out worse than the first. So the usual move- write it once, paste it into five places, call it "repurposing". Four of the five land flat.. so those four platforms must be dead for us.

They aren't dead. Each feed asks a different question about the post it gets handed. One caption answers one of those questions... if we're lucky.

Five Feeds, Five Questions

Every one of these platforms has published something about how it ranks, same as the YouTube home feed handing out one click per load.

LinkedIn asks how long we stayed. Its engineers wrote in 2020 that clicks and reactions are rare and noisy while dwell time is "always measurable"- then added a model predicting the chance a member scrolls past, cutting every post's score by "an amount proportional to the predicted P(skip) value". Scrolling past isn't neutral- it actively costs the post.

Threads asks what we'll do next. Meta's Transparency Center lists ten predictions its feed makes, among them how likely we are to reply and how likely we are to "scroll past a post rather than engage with it". No weights are published, so any Threads multiplier someone quotes is a guess.

X shows its working. The ranking code on GitHub predicts 24 separate actions per post (like, reply, repost, share by DM, mute, block, report...) and a params file holds the weights: reply 5.0, like 0.66, report -234 as of the October 16, 2026 sync. The README warns those multiply our own predicted chance of each action, not raw counts. The same code discounts a second post of ours in one feed load to 62.5% of its value, a third to 43.75%, down to a floor of 25%- the best one keeps full value.

A story doing the rounds says X went from publishing none of its weights to all of them in one week this summer. The repo history disagrees. Ten weights from the old heavy ranker have been public since April 2023. The new set landed in a single release on August 13, 2026, after three months with no commits. X's own help page still links the 2023 repo.

Instagram grades the post itself. Its Feed Recommendations card- posts from accounts we don't follow- lists "How informative a post is" as its first prediction, a signal "purely generated by LLM to assess the post's content quality".

TikTok first decides whether a post can compete at all. Anything under the "FYF Ineligible" standards won't be recommended in the For You feed. The list itself isn't on the eligibility page- it's scattered through "FYF Ineligible" sections across the whole Community Guidelines.

Rules in a File, Not a Prompt

A good prompt on Monday, a rushed one on Thursday missing the two clauses doing all the work... nobody types the same prompt twice. What holds is one plain text file per platform saying what it rewards and what it punishes- a "physics file", the laws of that one feed- with the agent writing against the file. The prompt shrinks to one disposable line:

Run the waterfall on idea.md, all five platforms.

Beside the five physics files sit idea.md (the idea and the angle, two lines) and voice.md (how we sound, words that aren't ours, what the agent may never write). Type an instruction twice and it belongs in a file- the same fix that pulled design work out of the generic look in Same Same, but Impeccable.

The folder idea.md two lines: the idea + the angle voice.md rules about you, written once physics/ linkedin.md x.md threads.md instagram.md tiktok.md 10 sections each, every line tagged + dated Agent reads every file before it writes one-line prompt LinkedIn- read to the end X- first line is the post Threads- leave room to reply Instagram- informative enough? TikTok- a script, not a caption wrong draft → new line in that file's corrective lines → same prompt again

All plain markdown- Claude Code runs it (paid plans only, not free claude.ai), so does any agent that reads a folder, or a plain chat window. Ten sections per file:

# Threads physics - verified 2026-10-19

## 1. My lane            who am I reaching here, and about what?
## 2. Ranking engine     what it rewards AND what it demotes
## 3. The one lever      one signal I design for + what I now do differently
## 4. Format             which formats it rewards, in what order
## 5. Shape of a post    opener rule / hard limit / band that performs
## 6. Cadence + timing   how often, how far apart, when
## 7. Harvested patterns what has actually won with my audience
## 8. Corrective lines   what the agent keeps getting wrong
## 9. Off-limits         what I won't do on this platform
## 10. Sources + dates   where every line came from, when it was checked

Sections 2, 3, 4, 5, 7 and 8 carry most of the weight. The demotion half of section 2 rarely gets written down, though suppression is what actually hurts an account.. nothing announces it.

Section 3 picks one signal and says what changes because of it, in a form that can reject a draft. "Focus on engagement" never says no to anything. "Is there a line here somebody could disagree with out loud?" does. The same lever on all five platforms is a belief about content, not a reading of each feed.

Official, Measured, Mine, Folklore

Every line carries one of four tags- official (the platform said it: docs, engineering blog, help centre, code), measured (a study stating both sample size and method), mine (measured on our own account) or folklore (heard everywhere, no source). A line that takes none of them stays out.

Folklore goes in on purpose. "A comment is worth fifteen likes on LinkedIn" has no LinkedIn source- its engineering posts give no numbers. The earliest trace we found is a 2024 blog crediting a paid report. Deleted, it creeps back from the next slide. Tagged, it stays recognisable, with only the documented direction kept: comments beat likes, by an unpublished amount.

The "800 to 1,000 characters" LinkedIn sweet spot gets pinned on a report of "1.8 million data points" whose summaries say 700-900 for image posts.. or 900-1,500, or 1,242-2,500, depending who's summarising. The 800-1,000 figure itself turns up in an AuthoredUp newsletter. No public method anywhere, so no band goes in- only LinkedIn's 3,000-character cap, tagged official, on its own line. The self-improvement world has the same disease- see the folklore numbers in Think Clearly to a Super You.

Date every file. A stale one is worse than an empty one, because the agent follows it confidently.

Teardown- Top Fifth vs Bottom Fifth

Section 7 can't be looked up- it's whether our readers want a story or an opinion, a question at the end or a full stop. Pull three to six months of posts from one platform (usually 20 to 40), our own account first. Write the ranking metric down before opening the data: the section 3 lever divided by reach. Raw likes mostly say which post is oldest. Take the top fifth and bottom fifth, set aside anything a big reshare carried, then:

Two groups of my posts on <platform> with their numbers.
TOP = highest <lever>/reach. BOTTOM = lowest.
Code every post on: opener type, length in characters, specific number (y/n),
format, ending (invites a response / closes the subject).
Report ONLY the dimensions where TOP and BOTTOM differ.
Where a dimension looks the same in both, say so and drop it.
No advice. No encouragement. Deltas only.

Without that last line the model signs off with a pep talk.. every single time. Whatever survives goes into section 7, dated and graded- high (three or more winners, mostly absent in the losers), medium, or hypothesis (one instance)- and the teardown re-runs quarterly.

When the Threads Post Comes Back as an X Post

Say the idea is that nobody quits the gym because the workout is hard- they quit on day three, walking in with no plan. The two drafts:

X:       Nobody quits the gym because the workout is hard. They quit because they walk in and don't know where to start.
Threads: Most people don't quit the gym over a hard workout. They quit because they walk in on day three with no plan.

Not one sentence carried over.. still the same post. On Threads it closes with nothing to reply to (replying is one of Meta's ten predictions), uses 109 of 500 characters, and has no scene a stranger could picture. Only the first problem is sourced. The other two are judgment, tagged mine, hypothesis, dated- out again if the next twenty posts don't back them.

Both files said the same three things- front-load the hook, concrete beats abstract, one idea per post- plus a 500 maximum on the Threads side. Given one idea and two files that agree, a model returns the shortest post satisfying both... an X post. The prompt was never the problem. If a line can be satisfied by doing nothing, it isn't finished: "500 maximum" needs a floor, "concrete" needs to say what concrete looks like here, endings need a rule. The fix goes in section 8, never in the post:

## 8. Corrective lines
[mine, hypothesis, 2026-10-19] Last line leaves a stranger something to add. If it closes the argument, it fails.
[mine, hypothesis, 2026-10-19] Use the room: 350-500 characters. My rule inside Meta's cap, not a second cap.
[mine, hypothesis, 2026-10-19] Put the reader somewhere. No scene = a thesis, not a post.

Same one-line prompt, nothing else changed:

Why do so many people stop going to the gym? Pretty sure it isn't the workout.

Day three. Through the door, thirty-odd machines, the two you actually know how to use both taken, no plan.. so you stand there five minutes working out where to even start, then leave. Going back next week is somehow harder.

That's not a discipline problem. That's five blank minutes.

What would have helped you- a written first month? Someone walking you through one session? Something else?

Checked line by line: the ending asks and offers two answers of its own, it runs 475 characters, and it drops the reader in a doorway with the only two familiar machines taken. Plan on the first three runs being bad- most of section 8 gets written in those. (Threads has also been rolling out 10,000-character text attachments since September 2025; the post body still stops at 500.)

What Stays Human

The ideas, our own stories and the replies never get generated. Ask for a post whose shape wants a personal opening and the agent will happily write one- fluent, in our voice, about something that never happened! The line in voice.md that stops it:

If you don't have the number or the story, write [FILL: what is missing]. Never estimate. Never source your own.

Start with one file, not five. The platform with some audience already, its six working sections plus cadence, every line tagged, the file dated. One idea through it, find the one thing wrong, add a line, run it again- and keep repeating those last two steps until the drafts stop failing before adding the next platform.

Which platform keeps flattening our posts, and what would the one lever be in its file? Got a folklore number that turned out to have a real source behind it? Leave a comment.. would love to see what ends up in those files.

Glossary

  • Physics file — a plain text file of rules for one platform: what it rewards, what it demotes, and how posts should be shaped there.
  • Dwell time — how long someone actually spends on a post once it's on screen.
  • P(skip) — LinkedIn's predicted probability that a member scrolls past an update; a higher prediction lowers the post's score.
  • Ranking signal / prediction — something a feed estimates about us and a post (will we reply, will we scroll past) and uses to order the feed.
  • Weight — the multiplier a ranker applies to a predicted probability before adding everything into one score.
  • For You feed (FYF) — TikTok's recommendation feed; X uses "For You" for its recommended timeline too.
  • LLM — large language model; the kind of AI that reads and writes text.
  • Agent — an AI tool that can read files and carry out steps on its own, not just chat.
  • Reach — how many people saw a post, measured the same way for every post being compared.
  • Teardown — comparing our best and worst posts to find what actually differs between them.
  • Folklore — a widely repeated claim with no source anyone can find.
  • Corrective line — a rule added to a physics file after the agent got something wrong once.
  • README — the front-page document of a code repository on GitHub.

Sources

Comments

Post a Comment

← Newer