How it works
Eight steps, and who does each one.
The whole product is one loop, run every week for every product. Each step below is marked with who does it: you, an AI model, or fixed code. Only two of the eight are a model.
01
It reads your site
Fixed code
You give it a product's address. It fetches the public pages and pulls out lines that contain something checkable: a number, a word like "fastest", or a promise.
Each one is kept with the exact line and page it came from. It leaves out what is not the site speaking for itself: quotations and testimonials, code samples, struck-out text, forms, the footer, sections that call themselves a demo or an example, and anything you mark with data-nosnippet. It still collects generously, because the next step is yours.
02
You confirm the facts
You
You tick the facts you stand behind. Nothing else can be cited. You can add a fact by hand when it is true but not on your site, and it is marked as hand-entered.
This is the only step that cannot be delegated. No model and no code ever confirms a fact for you.
03
It plans the week and writes the drafts
AI model
One request to an AI model plans three pieces, a social post, an email and a blog brief, and writes them. It is given your confirmed facts, your rules and a short note on which earlier pieces were clicked.
Every line it writes names the facts it rests on, and every piece comes with the reason it was planned. A run has cost about one cent in model fees.
04
Every claim is checked
Fixed code
Each draft is searched for three kinds of claim: figures, superlatives and promises. Any that does not rest on a confirmed fact gets the draft rejected, with the reason written down.
Your own rules are applied in the same pass: words you have banned, and disclosures you require. The check is code, so the same draft always gets the same verdict.
05
Someone approves
You
Drafts that pass wait in a queue. You approve or reject each one, or you let the reviewer do it for a channel once it has earned that.
The reviewer can pass, reject or send a piece to you. It cannot edit.
06
You publish, and it counts
You
Each approved piece gets its own tracked link. An email piece can be sent to you as a test, then prepared as a draft broadcast for a list you choose. You copy the piece out, or press send, yourself.
It does not post to social networks, and it never sends to a list on its own. Each time a link is followed a counter goes up. Nothing is stored about the person.
07
When a fact changes, pieces are recalled
Fixed code
When it reads your site again and a confirmed fact has changed, every piece that cited the old one is marked recalled. The new wording comes back to you as a candidate to confirm.
Facts you typed in by hand are not recalled this way, because there is no page to read again.
08
Next week starts from what happened
AI model
The click counts go into the next plan. You can switch a product to run every week without pressing the button.
Clicks are not sales, and a handful of clicks is not evidence. The results are used; whether they make the plans better is not something we have proven.
Who approves is your choice, per channel.
Manual
You approve every piece.
Where every new workspace starts.
Assisted
A reviewer agent approves routine pieces. You see anything it flags.
For channels whose drafts have stopped needing edits.
Autopilot
The reviewer agent approves everything it is allowed to. Each decision is in the ledger.
Your choice, per product and per channel.
Four things come to a person whatever you chose:
- The first piece on a new channel
- A claim resting on a fact no approved piece has used before
- Anything your rules mark as regulated
- Anything that spends money
What it does not do.
- It does not post to social networks. You copy approved pieces out.
- It does not catch a fact you confirmed that was wrong.
- It does not judge tone or taste. That is what your approvals are for.
- It does not find you an audience. It makes what you say to one consistent and checkable.