Skills & workflows

Skills do one job. Workflows connect them into a whole process.

A skill is a job an agent does the same way every time, such as calculating a suggested order from stock and sales. A workflow adds what belongs before and after it: the schedule on Monday morning, the stock levels from the warehouse system, the question to purchasing, the order to the supplier. You build both yourself or let the assistant build them.

A skill is a routine you would otherwise do yourself. You describe it once and the agent carries it out. It can contain many steps and building blocks: read documents, write records, use integrations, ask back. If something is off, you change it as often as you like.

How a workflow comes together

  1. 01

    Choose a trigger

    By hand, on a schedule, on an event in the platform or through a webhook from outside. A single skill also runs on a schedule without a workflow.

    • „Check stock levels every Monday at 06:00.“
    • „When an order comes in on Shopify, deduct the stock in the record.“
  2. 02

    Connect the steps

    Fetch, filter, branch, wait: these steps need no agent, cost no model calls and run the same way every time.

    • „Only items with less than 14 days of cover go on the list.“
    • „Orders above €5,000 go to management, all others to Jonas.“
  3. 03

    Approval where a person should decide

    The run pauses until a person says yes. Nothing goes to the supplier before that.

  4. 04

    Read every run and revert it

    If a run was wrong, you revert it. The records it wrote go back to their previous state.

  5. 05

    Have it built instead of building it

    Describe the process to the assistant or to your own agent over MCP. It creates the skill and the workflow as drafts, you publish.

    • „Build a workflow that calculates a suggested order every Monday and asks Jonas in Slack.“
    • „The suggested order should skip items with no sales in 90 days.“

Common questions

A skill defines what goes in, what comes out and which systems it may use. It has versions and belongs to the team. A run in September behaves like a run in August as long as nobody published a new version. A prompt is text that every person types differently.

The run stops. The log shows the step, its input and the error. You fix the cause, for example an expired connection, and start the run again. You can revert what the run has already written.

Yes. Integrations, filters, branches and loops need no model. An agent only comes in where something has to be understood or written: reading a document, drafting a reply, estimating demand.

Yes. Besides the skills that run in the platform, Adlass also stores skills for local agents. Your Claude Code loads them over MCP and uploads changes again. That way the whole team works with the same instructions.

Describe the task. The rest takes shape on the platform.

Start for free, create your first skill with the assistant, and see what a run looks like before you connect anything.

Start for free