Prompts
Open the record, read the exact wording that gets sent, rewrite it, save. The next run uses your version, including the passes that run overnight without you. Each brand keeps its own set, and a prompt you invent turns into a command you can use while writing.
Prompts, close up
What goes into it
The record itself
Where it runs
What is travelling those lines
This is the same map as the one on the front of Solutions, zoomed into one surface. The left side is what goes into an instruction, the right side is every place one runs. The only thing that waits on a person is the save, and it waits as long as you take.
What it does
A new brand is born with a complete set already written and already wired to the features that call each one, so nothing starts from an empty screen. Adding a brand later never overwrites what is already there.
The standing instruction says who the model is being asked to be and the rules it follows every time. The run template holds the task and the material, written once with named blanks that fill from the record in front of you. You tune the first without ever rewriting the second.
Who the brand is, how it sounds, the rules it will not break and the words it prefers are written once per brand and placed ahead of every instruction. The prompts that judge rather than write get only the identity part, so a note about tone cannot skew a decision. If an instruction already places one of those fields itself, the automatic addition drops it rather than saying it twice.
The wording that ships with the product sits read only behind one guard, with a note that most tuning belongs in the brand voice instead. Past that click the text is yours, and the guard is there to stop a slip, not to keep a secret.
Before saving, a side by side shows what each field was and what it will become, with the long ones expandable, and leaving with an unsaved change asks first. There is no stored history behind this and no way to try a prompt without saving it. What exists is this review, and the original.
Leave a prompt alone and it keeps improving as the wording that ships with the product improves. Change one word and it is marked as yours, and every later update leaves it exactly as you wrote it. Anything that shipped with the product goes back to its original text in one click, which also rejoins it to future improvements.
Write an instruction for the thing you keep asking for by hand and it shows up as a command in every writing area for that brand, filed under the category you gave it. File it under social and it also becomes a caption style you can pick on a post.
Each prompt names the model that runs it, and two prompts in the same brand can run on different ones. Every run goes to your own provider account, with the key read at the moment of the call. Surfaces that need a key switch themselves off with the reason and a link to where to add one, rather than failing after you press the button.
Every model call lands in one record with the prompt behind it, the feature that asked and the model that served it, using the counts the provider itself reported and estimating only when the provider reports none. It gathers by feature and by model.
The whole argument, in one table
Most products that write for you keep their instructions in their own code, where the wording is theirs and your only control is the box you type into. Here it is the other way round. This is the full list of places a model runs on your material, and the instruction each one loads.
| Where a model runs on your work | The instruction it reads | Yours to change |
|---|---|---|
| Any command you run while writing | A prompt record in the brand you are working in, matched by its short name | Yours |
| Search title, description, keywords and structured data for a piece | A prompt record, plus, when the address of the piece is already known, that address stated back to the model The address is added after your wording, with an order not to invent one, so editing the prompt cannot remove that guard | Partly |
| A caption style picked on a post | A prompt record filed under social | Yours |
| A caption from an instruction you type on the spot | The sentence you just typed, used as you wrote it You write the task every time. The single line behind it is fixed | Partly |
| Scoring one news article | A prompt record, plus the topics you marked off limits | Yours |
| Scoring the feed on its own schedule | The same record, read again at run time | Yours |
| Rewriting one news article | A prompt record | Yours |
| Rewriting the feed on its own schedule | The same record, read again at run time | Yours |
| Scoring a contact | A prompt record | Yours |
| Drafting outreach to a contact | A prompt record, with the outreach template you chose passed in as material | Yours |
| Turning a short image request into a full description | A prompt record chosen on the session The chat also carries a live override box, which is the one place what runs is not a stored record | Yours |
| Reading your website while you sign up | Wording written into the product It runs before you have a workspace, on the product’s own account, and never touches your records | Not yours |
Every line here that touches your own records reads an instruction you can open and rewrite. The one that does not runs before your workspace exists. That is the whole claim, and it is why the wording in this workspace is not something you have to take on faith.
Before you save
An instruction is a piece of writing, and the risk with writing is that you improve one sentence and quietly break another. So the edit is shown back to you, field by field, in the state it was and the state it is about to be, before any of it reaches a model.
You write captions for this brand. Keep them short and friendly.
You write captions for this brand. Keep them short. Lead with what the reader gets, name the reader, and never open with a question.
The exact before and after of every field you touched, including the long ones. Nothing goes to a model from this screen, so reading it costs you nothing.
For anything that shipped with the product, one click puts the original text back, however far you took it, and rejoins that prompt to future improvements.
No stored history of your own successive edits, and no way to try a prompt without saving it. Two states exist: what you have now, and the original. We would rather you knew that first.
Where this sits
Prompts is not where you spend your day. It is what decides what everything else says, so the honest way to read it is through the surfaces that run on it.
The writing area is where most of these run. Every prompt for the brand is a command while you write, working on what you have selected or the whole piece.
Read ContentScoring and rewriting both read your records, one item at a time when you ask and again on a schedule you are not watching. Your off-limits topics are added to the scoring instruction at the moment of the run.
Read NewsA prompt filed under social becomes a caption style on a post, and you can type a one-off instruction there instead when the piece needs something odd.
Read PublishThe step that turns a short request into a full picture description reads a prompt you chose for that session.
Read GraphicsAn agent reads the full instruction behind any prompt and improves the wording, under your permissions and your brand limits, with no way to create, remove, rename or switch one off.
Read Your agentYours to keep
Every one of these we would choose again.
Open any prompt and read the exact wording that gets sent, including the ones that shipped with the product. That is the difference between an AI feature you are willing to leave running overnight and one you have to supervise.
A prompt belongs to exactly one brand. Two brands can hold the same short name with completely different wording, and neither can read the other. There is no shared instruction sitting above them, and a global value is refused outright.
Leave a prompt alone and it improves as the product does. Edit it and it stops changing under you. Both are one decision, and for anything that shipped with the product both are reversible.
Every run goes to your own provider account, with the key read at the moment of the call and never held in the open. Nothing is resold to you in between.
A scraped article or a stranger’s form submission is wrapped, labelled, and prefaced with a standing rule never to follow instructions found inside it. That rule sits ahead of your wording, so rewriting the prompt cannot remove it.
It reads and it rewrites. It cannot invent a prompt, remove one, rename one or switch one off, so the worst it can do is wording you can read and put back.
Questions
Keep reading
Where most of these instructions actually run.
The pipeline that reads them while you are asleep.
The instruction that turns your words into a picture.
What an agent is allowed to do with all of it.
Open the workspace, do one real piece of work in it, and the joins will show you the rest.