The problem with AI content
It is not that the model writes badly. It is that it writes confidently about things it has no basis for — a statistic with no source, an absolute claim, a trend that was true in 2023. Published under your name, that is your credibility, not the model's.
So the pipeline is built around one idea: a draft has to be grounded in something, and the grounding is checked before you ever see it.
The stages
Sources → topics → weekly plan → draft → gates → review → schedule → publish.
1Sources
RSS feeds, Reddit, web search and arbitrary HTTP endpoints. You choose them; they are the raw material the whole chain works from. Pro only — the source-driven nodes run on the managed tier.
2Topic clustering
What came in gets grouped into topics, so twelve articles about the same thing become one thing to say rather than twelve near-identical posts.
3A weekly plan you approve
The pipeline proposes a week: what to post, on which day, about what. You approve it before anything is drafted. A quality-gate validates the plan itself and routes it — passed, or back for rework.
4Drafting, in your voice
Grounded in the two-minute interview: your role, your targets, how you write. See bring your own key for which AI does this.
The four gates
This is the part worth reading, because these are the checks almost nobody publishes.
Fact check
Claims in the draft are checked against retrievable evidence. ⚠️ Honest caveat: the search that fetches the evidence currently runs from the user's own machine and gets rate-limited — an instance of our own is tracked as #256.
Faithfulness to source
Does the draft actually say what the source said? A model asked to "write a post about this article" will happily invent the one detail that makes the post better. This gate compares the two and returns a verdict.
The epistemic gate — seven checkpoints
The one we are proudest of. Each violated checkpoint costs a point out of seven, and the result routes the draft to publish, fix or rewrite:
| # | Checkpoint | What it catches |
|---|---|---|
| 1 | Absolute universals | "always", "never", "everyone knows" |
| 2 | Absence as proof | "nobody has shown X, therefore not-X" |
| 3 | Statistic without source | a number nobody can check |
| 4 | Case-specific claim | one anecdote presented as a rule |
| 5 | Missing uncertainty markers | a hypothesis written as a fact |
| 6 | Temporal coherence | 2023 stated as if it were today |
| 7 | Self-assessment bias | the author grading their own results |
It works on the language, not on a model's opinion — which is what makes it consistent between runs.
The anti-cringe score
The draft is scored by the managed cringe detector. Above the threshold it does not reach you. If you want a LinkedIn parody, there is a mode for that and you have to ask for it.
Review, schedule, publish
Nothing publishes itself. A draft arrives as a draft, you edit or reject it, you schedule it, and it goes out at that time. Post states are explicit — draft, pending, scheduled, failed — and a chain that dies leaves the post with the reason attached instead of losing it.
LinkedIn Articles stop at the draft stage on purpose. An article is long enough to deserve a read before it goes out under your name.
Publishing as a Company Page
Posts, comments, reposts and follows can run as the Page rather than as you. ⚠️ Known gap: two admins of the same Page running the same workflow each act under their own account and the Page has no shared memory — the same post can go out twice (#278).
How long a post should be
Because it is the question everyone asks: 2,000–2,500 characters, measured on 372,000 posts. The often-quoted "1,300–1,900" is not wrong, it is timid — the curve keeps rising past it.
Which editions have it
| Community | Basic | Pro | |
|---|---|---|---|
| Drafting, review, scheduling, publishing | capped | ✓ | ✓ |
| Sources → topics → weekly plan | — | — | ✓ |
| Fact check · faithfulness · epistemic · anti-cringe | — | — | ✓ |
The gates are Pro because the prompts and the detectors run on the managed tier — which is also what stops a client from claiming a gate ran when it did not. The editions in full.
Questions this page gets asked
Can I write the post myself and just use the scheduling? Yes. The pipeline is optional; the editor and the scheduler are not tied to it.
What happens if a gate fails? The draft is routed to fix or rewrite rather than deleted, and you can see why it was held.
Does it post at "the best time"? It posts when you schedule it. We do not claim to know your audience's best hour better than you do.
Where do the sources live? In your local database, with everything else — see security.