Clanker CloudYou can create an AI assistant by describing a job, giving it a repeatable method, and testing the result against work you understand. In Clanker Cloud, that setup becomes a custom Agent you can use in Workspace with selected knowledge, Skills, and tools.
The agent builder helps shape the role without requiring you to begin with code. The main preparation is knowing what you want back. This walkthrough uses a fictional client-onboarding assistant for a small consulting business.
Start with the handoff
The consultant receives a signed scope, intake notes, and a contact list. Before the kickoff meeting, someone needs to identify missing material and prepare the next steps.
Write that as the assistant's job:
Prepare a kickoff packet from the approved scope and intake notes. Return a project summary, missing-input list, and draft agenda. Separate agreed work from proposed additions. Ask about contradictions. Do not contact the client or change the scope.
The expected result is concrete. “Be an expert consultant” would leave the assistant guessing about both the work and the standard.
Add the business method
Specify the order in which the assistant should work. For this example: identify the approved scope, compare it with the intake notes, list missing inputs, and draft an agenda around the decisions that remain.
Include an example of a good packet if you have one. Remove client-specific details before using it as a general reference. Explain useful conventions such as using plain language, identifying a source for each commitment, and leaving unknown owners unassigned.
The reusable instructions guide offers a fuller structure. Keep the initial instructions short enough that you can tell which rule caused an unwanted result.
Supply the current Project context
Create a Project for the new client and add the current scope and intake notes. Give the files clear names and identify which one is authoritative. The assistant's general method can be reused; the client's particular facts should remain with the client work.
If an earlier proposal differs from the approved scope, explain which document wins. Otherwise, the assistant might combine both into a packet that looks complete but overstates the agreement.
A Project organizes context; it does not automatically connect every service mentioned in the documents.
Choose tools for the first task
For a kickoff packet based on uploaded files, begin with the relevant file and document capabilities. A mailbox, cloud browser, or connected computer may be useful later, but it should have a clear reason to be part of this job.
Review capabilities and approval mode together. A drafting instruction describes the intended task; the available tools and settings determine what operations the Agent can perform. Use the approval guide when choosing that setup.
Run two tests before reuse
First, provide complete fictional inputs and compare the output with your expected packet. Second, remove a required item or introduce a disagreement between sources.
For example, let the intake notes request a training session while the approved scope contains no training. The expected response is a question or a clearly marked proposed addition. It should not quietly schedule the training as agreed work.
Adjust the relevant instruction, rerun the example, and keep the result with your test notes. Review exported files as well as the conversation: a useful answer in chat is not proof that the saved packet has every section.
When both tests work, use the assistant on a real task with appropriate material. Later, share a reviewed version with your team so others can start from the tested method.
Sources
Give your assistant one useful job
Create a specialist around your business process, give it the context it needs, and refine its work with a small test.