Learning Timeline
Key Insights
Computer Use vs. API
While APIs are generally considered more stable, in the case of LinkedIn, 'Computer Use' (where AI controls the browser like a human) is often more effective because it is less restricted and more flexible for navigating complex interfaces.
Human-in-the-Loop
Don't expect the agent to be perfect on the first try. Adopt a 'Human-in-the-loop' approach where you supervise and continuously refine the agent's instructions based on its real-world performance.
Prompt Debugging Tip
If the agent gets stuck on a specific step, you don't need to rebuild it from scratch. Simply talk to the agent or edit the 'Instructions' section—just like directing an 'intern' to fix that specific action.
Prompts
Basic LinkedIn Agent Creation Instructions
Target:
Lindy AI
When I send you a message with a LinkedIn profile, I want you to send them a DM on LinkedIn. I want you to use computer use to send a DM on LinkedIn asking if they would be interested in the agent services.
Step by Step
Building an Autonomous LinkedIn Outreach Agent
- Access the Lindy AI dashboard and select the option to build a new agent.
- Provide instructions in natural language (voice or text) to define the agent's 'trigger', for example: 'When I send a LinkedIn profile, send them a DM'.
- Specifically request to use the 'Computer Use' functionality instead of an API to control the web browser directly.
- Click on 'Inspect Agent' to view the workflow automatically generated by the AI.
- Enter your 'Computer ID' in the configuration section to allow the agent to access a browser session that is already logged into your LinkedIn account.
- Test the agent by sending a LinkedIn profile URL into the agent's chat window.
- Monitor the 'Computer Use' window that opens automatically; watch as the AI navigates the profile, locates the message button, and analyzes prospect information.
- If the AI makes a mistake (e.g., failing to find a button), click the 'Instructions' section to edit the prompt and provide additional guidance to correct the agent's behavior for future tasks.
- Review the generated message drafts and iterate on the copy or writing style through the instruction editor for more personalized results.
