Saarathi Academy in Kathmandu has released a structured four-week training guide designed for professionals who want to leverage AI without writing a single line of code. The curriculum explicitly rejects the common beginner instinct to start with exciting agent capabilities, instead enforcing a strict progression: safe habits, prompt engineering, simple automation, and finally, supervised agent usage. This approach addresses a critical infrastructure gap where non-technical users often deploy AI tools that amplify their own errors due to a lack of verification protocols.

Week One: Privacy and Verification Basics

The first week focuses entirely on configuration and trust. Trainees are instructed to adjust privacy settings on major platforms like ChatGPT, Gemini, and Claude to prevent their data from training model improvements. For instance, users must disable 'Keep Activity' in Gemini and 'Improve the model for everyone' in ChatGPT. The core lesson is that AI assistants predict text, not facts, so every date, name, and number must be verified externally. By the end of week one, students must identify three repetitive tasks and the time they consume, which serves as the foundation for later automation.

Weeks Two and Three: Prompts and Controlled Automation

Week two introduces structured prompting using a four-part framework: role, context, limit, and sample answer. Students practice on real-world tasks like translating Nepali to English or generating Google Sheets formulas. A significant operational note is included for Gemini users, as Google is migrating 'Gems' to 'skills,' rendering personal Gems read-only on November 17, 2026. Week three shifts to automation using Make.com, where students build a single flow—such as summarizing labeled emails into a spreadsheet—rather than attempting multiple integrations. This section emphasizes understanding triggers, actions, and webhooks, while warning that OAuth connections for Gmail are often the most friction-heavy step.

Week Four: Agents and Security Risks

Only after establishing verification habits do students encounter AI agents. The curriculum distinguishes agents from assistants by their ability to execute steps like clicking forms or opening pages. Saarathi highlights a specific security vulnerability known as indirect prompt injection, where hidden instructions in web content can hijack an agent's behavior. To mitigate this, trainees run their first agent tasks in isolated browser profiles with no sensitive credentials or payment methods linked. This supervised approach ensures that when agents inevitably fail, the damage is contained within a sandboxed environment rather than impacting live business operations.

Key Takeaways

  • Code is not required for effective AI usage; clear instruction and verification are.
  • Gemini 'Gems' become read-only on November 17, 2026, requiring prompt migration.
  • Start with one automation flow, not five, to manage complexity and credits.
  • Always place a human in the loop before any automated action sends, posts, or pays.
  • Indirect prompt injection is a primary risk for autonomous web agents.

The Bottom Line

Saarathi’s insistence on delaying agent deployment until week four is a necessary correction to the hype cycle; non-technical users must master verification and controlled automation first to avoid amplifying errors at scale.