Candu watches socials for what matters to your
Then it drafts a post, a video script, a message or an email, and waits for you to approve it.
You found out too late, and the opportunity to post has gone
How it works
Candu is built out of two things. A watcher says what to look for and how often. A skill says what to do with whatever turns up. Both are plain text you write, not software, so adding a subject to watch or a new kind of message to send is a piece of writing rather than a piece of engineering.
A watcher searches
Official accounts, the open web and X, on a schedule you set, anywhere from hourly to weekly. Anything seen on an earlier run is dropped, so the same story is never handled twice.
Every item is scored and taken apart
Each one is rated against your brief in plain English, with a one-line reason, and the facts are pulled out as fields: the areas named, when it starts, when it is expected to end, who issued it, and how severe it is.
Anything that does not clear the bar stops here
Below your threshold, or carrying a flag you said to refuse, and nothing further is spent on it.
Skills draft what makes the filter
One draft per channel, in your voice, saying only what the source said. A single item can produce several at once, a WhatsApp notice and a customer email from the same event.
You approve, and you send
Drafts wait in a queue, each beside the source it came from. Approve it, edit it and approve, or reject it with a reason. Approving marks a draft as fit to send. It does not send it. That last step stays with you, deliberately.
Ways people use it
A service business
Whose customers lose supply when a utility fails, and it did not cause the failure. A watcher reads the city and utility accounts alongside the open web every couple of days, counting an official notice for more than news coverage of the same event, and taking a resident's post when no official notice exists yet and the post names an area. One item then becomes two drafts at once: a WhatsApp notice for the suburbs affected, and an email to the customers in them. Both say only what the source said, and both wait for someone to approve them.
A marketing team
Who needs to publish fresh content across multiple themes. A watcher gathers information from the web and X on a custom schedule, and a skill drafts a video script from it, in the brand's voice and only from what the sources actually said. The script is text, so it goes wherever you send it: into a tool like HeyGen to have an AI avatar present it.
What Candu will not do
These are not switches with a sensible default. There is no configuration anywhere that turns them off, and changing one takes a reviewed change to the software itself.
Nothing publishes on its own
No channel, no exception, no scheduled send. There is no send button anywhere in the product: a person approves each draft, and a person sends it.
Private individuals and children are never written about
If the scoring step marks an item as identifying a private individual, or as involving anyone under 18, it never reaches a drafting skill. No watcher setting and no skill setting can override that.
No claim without a source
A drafting skill may only say what the source item says. It is instructed not to fill gaps from what it already knows, and to leave a field empty rather than guess at it. If a notice gives no restoration time, the draft does not invent one.
No certainty we do not have
Low confidence is recorded and shown next to the item, rather than smoothed over in the writing.
A style failure blocks
House style is checked on every draft. One that breaks it is regenerated once with the failure explained, and if it fails again it is held for a person and cannot be approved until it passes.
What it costs to run
Each watcher carries its own monthly budget and stops when it is spent, so the ceiling is something you set rather than something you discover. Every run records what it cost, and the queue reports the cost of each draft that was actually approved, which is the number worth watching.
What runs today, and what does not
Last updated: 2026-08-23
When the Resolver ships we will publish its precision figure, including if it is bad.