The Observer Pattern for Human Cognition Most personal AI projects focus on agents that act—sending emails, booking flights, or managing calendars. Šimon Podhajský argues for the opposite: a read-only "Observer" system named Fulan. By stripping away write permissions, we create a safe space for the AI to analyze "cognitive exhaust fumes"—the digital byproducts of our thoughts found in browser history, journals, and task managers. This system isn't a broken butler; it's a diagnostic tool for the human engine. Building the Fulan Architecture The system operates across three distinct zones. The sources remain read-only, ensuring the AI never contaminates the underlying data. Analysis occurs in the workspace, and insights land in a separate Obsidian vault. To implement this, Podhajský utilizes a Python script that orchestrates data retrieval and interfaces with the Anthropic API. ```python Conceptual logic for a Claude skill execution def run_weekly_reflection(): data = read_only_sources.get_all_activity() reflection = anthropic_client.generate_structured_output( prompt=PROMPTS['weekly_reflection'], context=data ) save_to_obsidian(reflection) ``` Cross-Source Magic and SQLite Integration The real power lies in cross-source signal detection. A standard CRM doesn't know what you're reading, and your browser doesn't know your contacts. By querying the Vivaldi SQLite database for browser history and matching it against a Clay CRM, Fulan identifies networking opportunities based on current interests. This requires "bash sorcery" on behalf of Claude to navigate local databases and map entities across silos. Security and the Lethal Triquetra Operating a system with this much personal data carries asymmetric risk. Podhajský references Simon Willison and the "lethal triquetra" of security: private data, untrusted content, and external communications. Even without write access, the mosaic effect—where small pieces of info form a devastatingly clear picture—remains a threat. The goal isn't perfect security, but a conscious examination of the risks you choose to carry.
bash
Programming Languages
Apr 2026 • 1 videos
High activity month for bash. AI Engineer among the most active voices, with 1 videos across 1 sources.
Apr 2026
- Apr 8, 2026