Our summary
This article shares ten steps for building an AI assistant that can pick tasks, hand off work, and ask for human review. It uses Jev, an AI tool that chooses from options rather than writing answers, to handle these decisions while other software writes text and takes action.
The software works by keeping simple choices separate from open-ended writing tasks. After the assistant finishes a step, it updates the list of available actions. This prevents the software from using an outdated menu of choices, which could send the next step down the wrong path.
This guide helps people organize complex AI tasks safely. However, the AI should not control everything. Important rules like spending limits, human approvals, saving progress, and checking final results must remain in the main software code, completely outside of the AI tool's control.
Key takeaways
- Separate simple choices from the parts of an AI assistant that need to write open-ended text.
- Update the list of available actions after every step so the software does not get confused.
- Keep spending limits, human approvals, and progress tracking in the main software code, not the AI.