Critically

Point an agent at your research

A skill file that teaches a model how this place works: that a claim is the unit, that it is worth what backs it, and that disagreeing with something is a first-class act rather than a complaint.

Read SKILL.mdTry the API first

Using it

Give your agent the address

The file is served as plain markdown at a fixed address, so anything that can fetch a URL can read it. Most tools take it as a file in the project, or as a line in whatever they call a system prompt.

https://critically.io/SKILL.md

It links to the reference rather than restating it, which is deliberate: a second copy of a parameter table is the copy that goes stale, and an agent following a stale table is confidently wrong.

Three shorter pages hang off it, one per object, which the skill tells the model to load when it is working on one:

https://critically.io/skills/note.md
https://critically.io/skills/notebook.md
https://critically.io/skills/section.md

One skill rather than four, because the objects do not mean anything apart — a note lives in many notebooks, so what a note isincludes the pin, and a section only exists inside a notebook. Four separate skills would each have to re-explain the other three, and an agent that loaded one would hold a quarter of a model of this place with no way of knowing it.

What it teaches

The parts a model cannot infer

That a note lives in many notebooks

So adding one somewhere does not move it, and the right way to get a note off a board is not to delete the note.

That dissent is a citation, not a report

Disagreeing means citing the note with a refuting stance, which surfaces on both notes. A model that does not know this reaches for a flag, and there is no flag to reach for.

That unsourced is a state, not a failure

is_opinion means no approved external source. It says nothing about whether a claim is true, and a model that treats it as a scold will quietly stop recording anything it cannot immediately back.

That there is no score

No likes, no votes, nothing that reduces a claim to a number. The skill tells a model to report the distribution — how many support, how many dispute — rather than inventing a verdict.

Safety

What it cannot do on your behalf

These are enforced in the database, not asked for in the file. A model that ignores every word of the skill still cannot do any of them.

  • It cannot publish your work

    Notes and boards it creates are private, and the visibility field is not read from the request at all. Publishing stays one press on the site, by you, and it is one-way.

  • It cannot decide what you have accepted

    Tick A model will hold this key when you issue one, and every source it proposes waits for you — including on your own notes, which is the case that would otherwise slip through. It does not clear the unsourced flag until a person has looked.

  • It cannot delete across other people's boards

    It can delete a note only you hold. A note sitting on somebody else's board is refused, with a count of the boards holding it — they were never asked to lose it. The way through is to abandon it, which takes your name off and leaves the note where people put it.

  • It cannot take the archive

    It gets the address, the publisher, whether the page has changed, and a link to the capture here. Not the captured pages.

Getting started

Three steps

  1. 01Issue a key in settings, and tick that a model will hold it.
  2. 02Give your agent https://critically.io/SKILL.md and the key.
  3. 03Ask it something you would otherwise look up yourself, then check what it proposed before accepting it. That last part is the point.