Stop asking AI for posts. Give it a world that can surprise you.

An agent can fill a calendar without learning anything. Give it a topic and it can turn that topic into twenty finished assets without producing one new event.

A better system changes one bounded world, records what actually follows, and lets the next episode inherit the result. What that system produces is a renewable source of events that neither you nor the agent can fully script in advance.

My position: the future will not have a shortage of content. It will have a shortage of things that actually happened.

I had automated the least scarce part

The pipeline could research a subject, draft a script, choose footage, generate a voice, add captions and render a short. It never had to wait for a real result. That convenience was the defect.

The machine had no reason to fail. If a claim was awkward, it softened the claim. If evidence was missing, it replaced evidence with explanation. If the topic was dull, it inflated the topic until it sounded historically important.

The output was usually competent, which made the problem harder to see. Bad content is easy to reject. Smooth, plausible, unnecessary content can fill a calendar for months. I had built a factory that could manufacture the appearance of having something to say.

The failure I observed was that throughput rose faster than specificity. The system measured completed assets. It did not measure whether a new fact, conflict or consequence had entered the world. A system can complete every production step without creating one new fact.

The production layer is becoming infrastructure

Production is becoming cheap, and events are not. The three images below are exact visual assets committed to real GitHub repositories. Editing, rendering and evaluation are becoming agent-addressable. The remaining strategic question is what the machine is being asked to discover.

Exact upstream files: the packaged bytes reproduce the Git blob identifiers recorded in the asset register.

AI writing and vibe-coded interfaces fail in the same way

They regress toward a competent middle. The result is rarely unusable. It is simply hard to believe that a particular person needed to make it. Writing and interface sameness are symptoms of the same missing thing: a particular decision.

The problem goes deeper than the word “delve”

Wikipedia’s field guide to signs of AI writing is useful because it notices deeper patterns: generic importance claims, superficial analysis, promotional tone, vague attribution and prose that could attach to almost any subject. The page also warns that no single pattern proves AI authorship.

Editing rule: replace importance with consequence, category words with objects, and unsupported significance with an event the reader can inspect.

AI UI has an accent too

Current critiques of vibe-coded products point to visual sameness, rounded-card repetition, muted gradients, polished default states and neglected error or edge states. Beige and rounded corners are not forbidden. The issue is that the interface has no evidence of a decision.

Distinctive design is evidence that somebody chose what matters, what can wait and what should happen when the happy path breaks.

A post is an output, and a world is a renewable source of events

The world does not need to be fictional. It can be a benchmark, a physical prototype, a simulated town, a drawer, a competition or a museum. It needs state, rules and a way to answer back.

A post generator starts with a subject and fabricates an output. A world engine starts with persistent state, changes one thing, then waits for reality, a simulation or an audience to answer back.

The difference is memory. A good episode leaves the world changed, and the next episode inherits that change. Viewers are no longer asked to remember a creator because of branding. They remember what happened.

world + intervention + uncertainty + proof + consequence + memory = episode

Persistent state, one intervention and permission to be wrong create renewable story pressure.

The Maintenance Monster turns invisible upkeep into plot

An automation claims to save six hours each week. Every hidden intervention, patch and exception feeds a physical creature. The episode ends when the ledger and the monster agree on what the automation is worth.

When every silent repair changes the creature, maintenance becomes plot instead of hidden labour. Different worlds should fail differently.

The equations cannot predict a hit, but they can stop me combining different failures into one fake score

Opening acceptance, retention, diffusion, audience memory and portfolio diversity are different processes. Each needs its own evidence and its own failure state. Do not compress attention, spread, return and creative diversity into one flattering score.

M1  \widehat a=(\ell c e)^{1/3}(1-m),\qquad \ell,c,e,m\in[0,1]
M2  \begin{aligned}S(t)&=\exp\!\left(-\int_0^t h(u)\,\mathrm du\right),\qquad h(u)\ge 0,\\[2pt]\mathbb E[\min(T,D)]&=\int_0^D S(t)\,\mathrm dt\end{aligned}
M3  \begin{aligned}N&=N_{\mathrm{ext}}+\sum_{g=1}^{G}N_g,\\[2pt]\mathbb E[N\mid R<1]&=\frac{\mu}{1-R}\end{aligned}
M4  \boldsymbol{\pi}_{t+1}=\boldsymbol{\pi}_tP,\qquad P_{ij}\ge0,\quad \sum_jP_{ij}=1
M5  \begin{aligned}\max_{\boldsymbol{x}\in\{0,1\}^{n}}\quad&\sum_{i=1}^{n}q_ix_i\\[2pt]\text{subject to}\quad&\sum_{i=1}^{n}c_ix_i\le B,\\&\sum_{i\in\mathcal C_m}x_i\ge1\quad\forall m,\\&\sum_{i\in\mathcal W_r}x_i\le\rho\sum_{i=1}^{n}x_i\quad\forall r\end{aligned}

Test which world makes relevant people return

Do not search for the viral hook. The first pilot should compare different plot engines. Cosmetic hook variants can wait until there is evidence that the underlying world deserves another episode. Compare content worlds over repeated episodes, not isolated hooks over one weekend.

The primary outcome is qualified subsequent-episode consumption: did a relevant viewer voluntarily consume another episode from the same world within the declared window? Four guardrails sit beside it: evidence completeness, trust incidents, production burden and portfolio concentration.

One hit does not crown a format. The system should earn confidence through repeated episodes, comparable cohorts and a reasoned account of what changed.

Automate the labour, and keep public authority bounded

An agent can research, propose experiments, write, render, test and package a release. That does not entitle it to decide that a claim is true, a person is safe to depict or a piece is worth publishing.

I still want automation. I want agents to do the dull research, assemble the evidence pack, render the variants, run the checks and remember what happened last time.

I no longer want the agent to begin with “make content”. I want it to begin with a world, a rule and a question whose answer is not yet known. Distribution, for me, is accumulated memory rather than accumulated output.

For coding agents: give the agent a world, not a topic

The useful instruction is a bounded operating contract that forces state, uncertainty, proof and memory before production, rather than “make content about AI”.

Before creating content:
1. Name the persistent world state.
2. Change one thing on purpose.
3. State what remains genuinely unknown.
4. Lock the evidence needed before the run.
5. Keep an unflattering result.
6. Write the state change back into the world.
7. Ask for human approval before public release.