Building your own AI assistant in Claude

Tap a module to see its contents

1. 1Choosing the task
   • –Participants describe their main work tasks to the AI and answer clarifying questions.
   • –We compare how often each task recurs, how available the information it needs is and how verifiable the result is.
   • –We choose one clearly defined task whose solution can be finished within the day.
2. 2Defining the result and the way of working
   • –We describe what information the work starts from, what steps need to be taken and what the result must look like.
   • –We add an example of a good result and define when the AI must ask for more information or for human approval.
3. 3Connecting information sources
   • –We choose the files and services the task needs and use approved data connections (Connectors) to access them.
   • –We check whether Claude finds the right material and what actions the connection allows.
   • –We distinguish between reading material and changing it.
4. 4Creating the project and project instructions
   • –We create a Project for the chosen work task and add the materials it needs.
   • –In the project instructions, we write the goal of the task, the requirements for the result and the limits of what it may do.
   • –We define when the assistant must ask for more information or for human approval.
5. 5First run of the task
   • –We have Claude carry out the task from start to finish based on the project instructions and materials.
   • –Along the way, the trainer shows the techniques needed, such as searching for information, reading files or formatting the result.
   • –We refine the instructions where the result needs improvement.
6. 6Creating reusable work instructions
   • –We turn the way of working that works into one or more work instructions (Skills).
   • –Project instructions set the goal and limits of the work; a reusable work instruction describes the steps of a specific activity.
   • –We add the work instruction to Claude and test using it in the project’s next task.
7. 7Testing and improving
   • –We test the solution with a new example and in a situation where some information is missing.
   • –We check whether the result meets the agreed requirements and whether the AI stays within the limits of its task.
   • –We improve the instructions and try again.
8. 8Scheduling and further use
   • –The trainer shows how to create, change and pause a scheduled task in Cowork.
   • –If the chosen task needs to be done every day, participants add the necessary instructions, materials and connections to the scheduled task and test how it works.
   • –We check separately that the scheduled task receives the background information it needs.
   • –If there is no daily need, participants do not schedule their task.
   • –We write down how to run the solution and how to check its first results.