News manager
Point it at the feeds you would otherwise skim every morning. It fetches them on a clock, opens the article behind the headline, scores each one against a paragraph you wrote about your own audience, and drafts the ones that pass into your content table. Repeats are dropped on the way in, and nothing publishes itself.
News, close up
You set it up
Where it is judged
What comes out
What is travelling those lines
The same map as the one on the front of Solutions, zoomed into one surface. The two pieces in the middle with your handwriting on them are the scoring instruction and the writing instruction. Everything on the left exists to get articles to them, and everything on the right is what they hand back.
What it does
You write, in your own words, what your audience cares about and what they do not, and that text is what every article is judged against. It is held per brand, so two brands in the same workspace can disagree about what matters, and the answer comes back as a score out of ten, a decision, a topic and a reason you can read.
The moment you change either instruction it is marked as yours. When the shipped starting text improves, the update walks past your version and leaves it alone. Nobody quietly replaces the sentence you tuned, and an agent editing it for you goes through the same record.
Most feeds hand over a headline and two lines. A separate step opens the article page and stores the readable text off it, and both later steps prefer that text when it is there. It needs no key, and it never changes where an article sits, so a page that refuses to be read costs you nothing.
A fingerprint of the article address is the key on the row, so re-polling a feed adds nothing new. The tidy-up keeps that fingerprint even when it archives the article, which is why something you cleared away in one month cannot come back and be judged again in the next.
Every poll records what happened. Consecutive failures build a streak, and at the limit the source switches itself off and says so, with one click to bring it back and clear the streak. Peeking at a feed from the side panel is a real poll, so the health you are looking at is current rather than remembered.
Mark a topic as never publish about this and a hard rejection naming it is bolted onto your instruction at the moment it runs. It beats the score, so an article mostly about that subject is turned down with the topic named as the reason, and you never had to rewrite your own paragraph to say it.
Fetching and reading articles are free and run over everything. Judging and writing cost you money, so they run on the rows you ticked, and each button tells you how many of your ticked rows it actually applies to. Two hundred articles arriving is not two hundred calls.
Every judging and every writing call is metered against the article it was for, including the one retry a malformed answer earns. It shows on the row and in the totals, so the question of what a busy week cost has an answer rather than an estimate.
A row that fails keeps a plain reason and a count of how many times it has failed, both cleared by a later success. Retrying re-runs the step that actually failed. Anything stranded halfway by a crash is picked back up at the start of the next fetch, and a bulk run reports its failures grouped by reason rather than as a number.
As the backlog grows, the tidy-up archives articles that were never drafted and have gone stale, and strips the heavy stored payload off the ones you are finished with. Both are updates. Nothing is deleted, and anything you have drafted is untouched.
One article, all the way through
Below is a single article making the whole trip, and what the workspace is holding after each step. The feed name, the headline and the counts are made up for the walk. Everything else is what actually happens, in the order it happens.
01No key needed
Paste the link. The name and the key are suggested from the address, you pick the topic to file it under, and you can preview the live feed in the side panel before you commit to it. Each item in that preview is badged as new or already known, so you switch the source on once you have seen it parse.
What the workspace has now
02No key needed
You picked a frequency once and left it. It polls every switched-on source in the workspace, writes down what each one did, and files what came back. Nothing sets this off, it is a plain scheduled job on a timer, and it appears in Automations with the rest of them.
What the poll wrote down
03No key needed
The feed gave you a headline and two lines. This step opens the article page and stores the readable body text off it, with the time it did so. It costs nothing and it does not move the article along, so a page that refuses to be read leaves everything else exactly as it was.
What the row carries now
04Runs on your key
Not a relevance slider and not a keyword list. The instruction is a paragraph you wrote about what your audience cares about, held for this brand, read at the moment of the call together with any topic you have ruled out. What it returns lands on the row in words, whether the article passed or not.
What it decided, and what it cost
05Runs on your key
Your writing instruction decides the shape: structure, length, tone, what to preserve, what to attribute. It reads the stored article text rather than the teaser, writes in the language this brand is set to, and refuses rather than guessing if the brand has no language set.
What came out
06Yours
The piece is in your content table, in planning, waiting. Nothing on this surface publishes, whether you clicked the run or a clock did. You open it in Content, edit what you want to edit, and publishing is a separate act you perform there, on the piece.
What is left for you
Where this sits
A feed is only worth watching if what comes out of it lands somewhere you already work. Each of these has a page of its own.
Both instructions live here as records you open by name and rewrite. The scoring one decides what survives the feed, the writing one decides what a piece looks like, and changing either marks it as yours.
Read PromptsDrafts land here as news posts in planning, with a readable address, their tags, their search metadata and a link back to the article they came from. Same table, same editor, same statuses as everything else you write.
Read ContentThe separate act at the end. Once you have read and edited the piece, sending it out is something you do to the piece, not something this surface can do while your back is turned.
Read PublishingAn agent registers and updates sources, lists what arrived, reads one, and rewrites both instructions. It does not run the steps. It curates the rules, not the runs.
Read Your agentYours to keep
Every one of these we would choose again.
The one thing a tool like this usually decides for you is the thing you own here. Relevance is your paragraph, per brand, and it is the same text whether the run was one you clicked or one that happened while you were asleep.
Every piece that comes out of here is created in planning, and no path off this surface moves it further. Publishing is a separate act, on the piece, by a person. That is a decision we made, not a gap we have not got to.
The judging and the writing run on a key you connected, and every call is attached to the article it was for. Nothing is resold to you, and nothing is spent without a row you can point at afterwards.
Fetching and reading articles never need a key. A workspace with no key connected still gathers its sources, drops the repeats and stores the article text, and the two steps that do cost money are switched off in plain sight rather than failing halfway through.
Drafts come out in the language this brand is set to, read from the brand record. A brand with no language set gets a clear refusal instead of a guess, which is the failure worth having.
Topics are your own list rather than a fixed menu. They file the sources, they are what the scoring hands back, and a topic in use refuses to be removed while anything is still filed under it.
Questions
Keep reading
Where both instructions live, and where you rewrite them.
Where the drafts land and where you edit them.
The separate act that sends a finished piece out.
What an agent may and may not do with all of this.
Open the workspace, do one real piece of work in it, and the joins will show you the rest.