PRD: Project Jennifer (Autonomous Personal Assistant)

1. Vision & Core Philosophy

The goal was to build a Digital Site Reliability Engineer (SRE) and Personal Assistant capable of intent-based execution rather than pre-defined logic.

2. Technical Requirements

• Persistence: Must utilise a Markdown-based memory system (MEMORY.md) to maintain state across sessions.
• Agency: Must have shell access to an Ubuntu server to autonomously build, deploy, and manage Docker containers.
• Interfaces: Integration with Telegram for mobile interaction and a local Web UI via Tailscale.

3. Security & Governance Requirements (Post-Audit)

• Firewall: UFW must implement a deny-by-default policy, allowing only Tailscale and LAN-specific traffic for dashboards.
• Access Control: SSH must be hardened to Ed25519 keys only; root login and password authentication must be explicitly disabled.
• Integrity: A weekly automated scan must validate SHA256 signatures of all third-party “skills” to prevent supply-chain attacks.
• Privacy: All model interactions must adhere to zero-data-training policies; models that train on API usage are strictly prohibited.

4. Success Metrics & Retrospective

• Deployment Efficiency: Goal: Build and deploy internal apps in <3 hours. Result: Success.
• System Reliability: Goal: Proactive notification of network issues within 5 minutes. Result: Success.
• Economic Sustainability: Goal: Monthly API costs <£25. Result: Failed. Background polling for 24/7 agency created an unsustainable “Agency Tax”.

5. Final Specification: The Tooling Model

The project concludes that 24/7 autonomous agency is currently a “luxury” cost. The final requirement is a Hybrid Model:

• Monitoring: Deterministic Cron scripts (Cost: $0).
• Execution: Claude Code via manual trigger (Cost: existing subscription).