If you want AI training in Kathmandu and you do not write code, learn it in this order: safe habits, then prompts and spreadsheets, then one automation, then one AI agent. Agents come last, because an agent repeats your mistakes faster than you can catch them.
Most beginners want to start with the exciting part. We think that is the wrong way round. Each week below depends on the habit you built the week before.
Do you need to code to learn AI in Kathmandu?
No. Office staff, shop owners, teachers and students can use AI well without Python. You need a laptop with Chrome, a Gmail account and the patience to check answers.
Code matters if you want to build AI products. Using AI at work needs different skills: clear instructions, careful checking, and knowing what never to paste.
Week 1: set up safely before you ask anything
Open free accounts on two assistants, for example ChatGPT and Gemini. Before your first real question, open the privacy settings.
In Gemini, turning off Keep Activity means your future chats are not used to train Google's models. Google still keeps them for 72 hours so Gemini can reply. In Claude, the switch is called Model Improvement. In ChatGPT, it is "Improve the model for everyone", under Settings, then Data controls.
Then learn the one fact that changes how you use these tools. A chat assistant predicts the next word. It can give a wrong date or a made-up rule in a confident voice. So from day one, check every number, name and date outside the chat.
End the week with a list of three boring tasks you repeat, and the minutes each one takes. That list decides what you build in week 3.
Week 2: prompts, writing and spreadsheets
A good prompt has four parts: a role, the context, a limit, and a sample answer the AI can copy. "Write a reply to a customer" gets a weak answer. "You are a polite shop assistant. A customer asks if the red kurta in size M is back. Reply in under 60 words, in simple Nepali. Here is a reply we liked last week" gets a useful one.
Practise on real work: emails, price quotes, Nepali to English translation, a summary of a long PDF. Keep your best prompts in one Google Sheet, so you stop writing them again.
Spend real time on spreadsheets. Ask the AI to write a Google Sheets formula, then test it on three rows you can check by hand. If one row is wrong, paste that row back and ask again.
A note if you use Gemini. Google is moving Gems, its saved custom assistants, to a new feature called skills. On personal accounts, Gems become read only on 17 November 2026 and move to skills by themselves. Keep a copy of your best prompts in your own sheet anyway.
Week 3: one automation, built on purpose
An automation is a task the computer repeats for you. Learn four words first. A trigger starts it. An action is what happens next. A webhook is how one app rings another, and JSON is the labelled parcel of data they pass along.
Build one small flow, not five. A good first one: you put a Gmail label on an email, the AI writes a two line summary, and the summary lands as a new row in Google Sheets. The free plan of Make.com is enough for this.
Expect the Gmail connection to be the slowest step. A personal Gmail needs a Google Cloud sign in key, called an OAuth client, before Make.com can read it.
Run the flow by hand while you learn. In Make.com, a scenario set to "on demand" waits for the Run once button and does not run by itself. That saves your free credits, and a broken flow cannot run all night.
Why not start here? The AI step inside the flow is the same chat you used in weeks 1 and 2. If you never learned to check its answers, the flow now writes wrong rows into your sheet with nobody looking.
Week 4: a human check, then your first agent
Before any flow sends, posts or pays, put a person in the middle. The AI drafts, you approve, then it sends. Break your week 3 flow on purpose with bad input, read the error log, and fix it.
Only now try an AI agent. An assistant answers you. An agent goes and does the steps: it opens pages, clicks and fills in forms.
Agents carry a risk most beginners never hear about. A web page or a file can hide instructions written for the AI, and the agent may follow them. OWASP, a security group, calls this indirect prompt injection. So run your first errand in a separate browser profile with nothing important signed in. Use public pages only, with no passwords and no payments, and watch every step.
How Saarathi Academy teaches this order
Saarathi Academy in Old Baneshwor, Kathmandu teaches this order in its No-Code AI, Agents and Automation course. It runs five weeks, two hours a day, Monday to Friday, in person and online, with up to 10 students in the classroom.
The first two full weeks are safe use and everyday AI, before any automation. You build one automation fully in Make.com, with a human check on every step that sends, posts or pays. You run one AI agent errand under supervision. You write no code at any point, and the course ends with a certification exam taken online at home.
If you want the paths that do need code, our guide to AI and data science training in Nepal covers them.
What to try tonight
Open a Google Sheet and type ten rows of real numbers, like a week of sales, with a column that says "cash" or "QR". Ask ChatGPT or Gemini for a formula that adds up only the cash rows. It will likely give you something like =SUMIF(C2:C11, "cash", B2:B11). Paste it in, then add up three cash rows by hand and compare.
Then open Gemini's activity settings and decide what you want Google to keep. For Make.com, read how on demand scheduling works before you build anything.
The course page has the details of the No-Code AI, Agents and Automation course, and the full week by week syllabus is public.
Top comments (0)