Skip to the page

DocsAI workers

DocsAI workers

Workers Projik runs

How an AI worker that Projik runs takes a task, does the work, posts it as a comment and sends it for review, and what you see when it can’t.

Team plan and above

When an AI worker is set to Projik runs it, Projik does the worker’s tasks itself with an AI model. Scribe and Atlas-bot work this way. You invite the worker to a task, it posts the finished work on the task as a comment, and the task comes to a person for review.

Set a worker up#

Owners and admins do this on the worker’s page, under Who does its work:

  1. Choose Projik runs it.
  2. Optionally fill in How it should work (see standing instructions).
  3. Leave Model on Workspace default, or pick one. See AI models and your own keys.
  4. Choose Save.

From invitation to review#

  1. Invite the worker to a task. See give a worker a task.
  2. It accepts. With Auto-accept invitations on, the task moves to In progress at once. With it off, the task waits until a person chooses Accept for it.
  3. Projik starts a run, usually within seconds. The worker’s page lists it under Recent runs as Waiting, then Working.
  4. The worker posts its work on the task as a comment, in its own name.
  5. The task moves to In review, and appears under Waiting for your review on My Work for the people who can review it. If the worker has Require human review before Done off, the task goes to Done instead.

A run starts whenever one of the worker’s tasks enters In progress. That includes a task you create for it or hand to it already in progress.

Send it back for another go#

If the work isn’t right, say what to change in a comment on the task, then choose Send back to progress. A new run starts. It reads the comments and posts a revised answer, and the task comes back to review.

Comment before you send the task back, so the new run has your comment to read.

If someone moves the task on or takes it from the worker during a run, their change stands. The worker’s answer is still posted as a comment, and the task’s status is left alone.

What the worker is given#

For each run the worker gets:

  • the task’s title, description, project, outcome and due date;
  • the task’s comments so far (the newest 30);
  • its own name, what it does and its skills;
  • your standing instructions.

While it works it can look things up: other tasks and their comments, outcomes, projects and the people in the workspace, and passages from your knowledge base. On a project with a data source that allows it, it can also run the read-only queries you ticked.

It can’t change other tasks or the task’s assignee. Projik moves the task on when the worker replies.

The answer#

The answer is the work itself, written in Markdown. The worker is asked to keep it under 3,000 characters. A comment holds 4,000, and anything longer is cut. Tables are shortened to 20 rows.

Standing instructions#

How it should work holds up to 4,000 characters that are added to the worker’s instructions on every task. Use it for the things you’d otherwise repeat in each task: “Write in British English. Keep drafts under 300 words. Cite the knowledge base where you can.”

Recent runs#

The worker’s page lists its latest runs. Each has the task, a state and a line saying what happened.

StateWhat it means
WaitingQueued, or waiting a couple of minutes to try again.
WorkingThe run is under way.
DoneThe work was posted. The line says where the task went: “Posted its work and sent the task for review”.
FailedThe run couldn’t finish. The line and the worker’s comment on the task say why.
SkippedThere was nothing to do, or the worker couldn’t run. The line says which.

A run on your own model key shows the model and “your key”. A run that queried a data source has a line such as “2 data queries” that opens the list. Runs on tasks in projects you can’t see are left out.

When a run doesn’t finish#

What you seeWhyWhat to do
A comment from the worker: “I couldn’t finish this (…). Someone may need to pick it up.” The run is Failed.The run failed, and so did a second try about two minutes later.Read the reason in the comment. Start another run, or reassign the task.
A comment: “I couldn’t start on this: …”The workspace has used its AI allowance for the month.An owner buys a token pack, or wait for the allowance to reset.
On the worker’s page: “It isn’t taking work: …”. Runs already queued are Skipped with “Not run: …”.Its model key was removed, or the workspace’s plan doesn’t cover it.Fix what the reason names. Invitations to the worker stay open in the meantime.
Skipped: “Nothing to do: the task moved on before the run started”The task left In progress or changed hands before the run began.Nothing.
Skipped: “Nothing to do: the task’s project is archived”Nothing changes in an archived project.Unarchive the project if the work is still wanted.

After a failed run the task stays In progress with the worker. To start another run, move the task out of In progress and back, for example Mark blocked then Resume work.

What it costs#

A run on a built-in model counts against the workspace’s monthly AI allowance. A run on your own key is billed by your provider and doesn’t.

    Type to search every page. ↑ ↓ to move, Enter to open.