Some workflow steps should not run for every work item. Others should run, but should not proceed to the next step until the work item is in a certain state.

Workflow Options on a step lets you declare both rules with the same condition builder used elsewhere in Thunk (for example in tests):

Thunk checks these conditions directly against the work item's values — not by AI judgment in directions. A Step Finish Condition can also include optional AI checks for things a condition cannot express, such as whether an uploaded document is the right kind of document (see Add AI checks to a finish condition below). Both are different from a tool’s Availability condition, which only limits when a tool can be used inside a step that still runs.

Use this when the decision is based on work-item property values you can express as a condition — for example:

Prefer this over writing “skip this step if …” in directions when you want a deterministic skip (same property values → same skip/run behavior, no AI call on skipped work items).

Conditions can use most properties on the work item (including columns that are not bound as inputs for this step). Some system bookkeeping columns are not offered in the picker. If the work item has no properties yet, the panel shows that no properties are available.

When properties are available:

The panel shows a short summary of the saved condition. Use Edit Condition to change it, or Clear to remove it so every work item runs the step again.

For a work item that does not meet the run condition:

When a condition is set and the work item does match, the step runs normally. The step chat may also show an agent message: Run condition passed; running this step., with the same kind of per-condition breakdown.

Below the run-condition controls is a separate toggle: Skip this step for test work items.

Turn this on when a step should not run during automated test re-runs — typically side-effecting steps such as sending email or writing to an external system. On a re-run of a work item that carries test assertions, that step is skipped and the work item advances; the step chat records an agent message: Step skipped because this is a test row. (“Test row” here means a test work item.)

You can use a run condition and the test toggle on the same step. The test skip applies when the work item is a test work item being re-run; the run condition applies to whether the work item’s properties match your condition.

Tip: on a test re-run, the step you start from may still execute; the toggle is most useful for later steps in the plan that should stay quiet under test.

Use this when the step should run, but the workflow should not proceed to the next step until the work item reaches a required state — for example, only proceed when ReviewState is Complete.

This uses the same Add Condition dialog as the run condition. It is a hard stop on automatic proceed (similar in spirit to requiring a human to finish the step). A person can still finish the step manually from the step controls.

When the agent is ready to finish the step and move on:

Some requirements cannot be written as a column / operator / value condition. For example, an expense workflow expects a receipt in its Receipt file property, but someone uploads a washing-machine manual instead. The file is present, so a condition on the property passes, yet the work item is wrong.

AI checks cover this. Each one is something an AI model must judge true before the step can proceed.

Quickest option — check outputs against their descriptions. Under Step Finish Condition, turn on Check that outputs match their property descriptions. That's it: no dialog. An AI model compares each of the step's output values with its property's type and description. File and image properties are checked by their content, so a manual uploaded where the description says "a PDF of an expense receipt" is caught. This works best when your output properties have clear descriptions.

Your own statements. For anything the descriptions don't say:

Clear removes your statements; it does not turn off the descriptions check, which has its own switch.

AI checks run when the agent tries to finish the step, after any finish condition above. If every check passes, the workflow can proceed as usual. If any check fails, or the model cannot confirm it, the step stays open exactly as it does for an unmet finish condition. The conclude message shows Did not finish step because finish conditions were not met followed by one line per check, for example - AI check "The Receipt is an expense receipt, not some other document" → failed — the document is a washing machine manual.

Things to know: