How it works
Four stages, running continuously.
The interesting part is not any single generation. It is what happens between them.
Stage one
Extraction — it reads you before it writes anything.
Onboarding takes about ten minutes. You describe your role, your audience, what you want to be known for, and the topics you have standing to talk about. Then you paste writing samples — past posts, newsletter sections, long comments, anything that sounds like you when you are not performing.
From that, the system extracts named traits and shows them to you. Nothing about the read is hidden inside a prompt. If it says your register is formal and you know it is not, you change that trait and the change takes effect immediately.
Stage two
Grounded generation — your material, not the model's guesses.
Before a draft is written, the system retrieves the most relevant pieces of your knowledge base and any research you have attached, using hybrid keyword and semantic search. Those excerpts are handed to the model as evidence.
The consequence is specificity. The draft can reference the actual framework in your deck or the actual number in your report, instead of the plausible-sounding alternative a general model would produce.
Retrieved content is always treated as data. If a document contains text that looks like an instruction, it is not followed.
Adoption stalls after the pilot in most enterprise rollouts
Linked source: industry survey
Teams cite integration cost as the top blocker
Linked source: vendor report
A fact without a source link does not render. That is a hard rule, not a preference.
Stage three
Edits — the feedback that is actually honest.
Diffed, not just saved
When you save, the system compares your version against what it generated and classifies the change: shortened, de-formalised, story added, call-to-action removed.
Patterns, not one-offs
A single edit means little. A change you make eight times out of ten becomes a rule the next draft respects.
Weighted by recency
Older signals decay, so the profile tracks how you write now rather than how you wrote a year ago.
Stage four
Performance — the audience gets a vote.
Once your LinkedIn account is connected, published posts are matched back to the drafts that produced them and their metrics are pulled in on a schedule. The system looks for structural patterns: which hook types, lengths, formats and pillars outperform your own baseline.
Those patterns become suggestions rather than rules. You are told what the data appears to say and left to decide whether it is worth acting on, because a metric is a proxy for what you care about, not the thing itself.
See how close it gets on your first try.
Paste a couple of posts you have already written. You will know within a minute whether this sounds like you.