Online test series: CUET, IPMAT, CAT, SAT Buy test packs
Built with ChatGPT Codex

How I built my AI secretary with ChatGPT Codex

You just watched a bot block my class on the calendar and my Topmate slot disappear on its own. Here is how that happened, the stack behind it, the mistakes that nearly broke it, and the exact prompt to build your first version tonight.

I did not write a single line of code. I built it with ChatGPT Codex. My job was the part AI cannot do for you: knowing exactly what problem to solve, and checking every step before trusting it.

2 daysof building
Rs. 0a month to run
6build stages
70automated checks
By Anshul Wadhwa, Career Counsellor and Founder, AptiGuide · October 2026
Disclosure: ChatGPT Pro access was provided through an agency campaign.

The problem was never "I need an app"

I teach classes, counsel students and families, and take paid sessions on Topmate. Three things kept going wrong:

Before building anything, I checked what already exists. Google Calendar, task apps and paid schedulers each solve one piece. None of them remembers a promise I made to a specific person and keeps reminding me until it is done, without spamming me. That gap is the only thing I actually built. Everything else is free software I connected.

The stack, and why each piece

Tap a card to see why I chose it over the obvious alternative.

The process, in six steps

This is the part most people skip, and it is why most AI projects die after the demo.

1Write the problem, not the solution

I did not ask AI for "a scheduling app". I wrote down what actually goes wrong in my week: rotating classes, forgotten promises, a booking page that does not know when I am busy, a team that has to ask me. A precise problem is what makes the AI's output precise.

2Test the riskiest assumption before writing any code

The whole idea depended on one thing: would Topmate hide a slot if I blocked that time on my calendar? I tested it by hand in twenty minutes, in a private browser window. It worked, so the plan was safe. If it had not, I would have redesigned before wasting two days.

This one habit saves more time than any prompt. Find the single assumption that, if wrong, kills the project, and test it first.
3Plan the design with AI, decide it yourself

I used AI to plan the stack, the data it needs to remember, and the rules it must follow. Then I made the calls: which account it runs on, that it can never send email, that my team sees busy times but never details, that it never messages me after 10 pm.

4Build in small stages with ChatGPT Codex

Six stages, each one working before the next began:

  • A bot that only replies "I heard you" (proves the connection)
  • Calendar: read my day, add blocks, cancel, with Undo
  • Promises, tasks and things I am waiting on
  • Reminders that escalate, with a hard daily limit
  • Class slots declared in two taps, which also block Topmate
  • My team: busy times only, requests I approve with one tap
5Check everything the AI writes

AI-written code looks finished before it is. Every stage was reviewed and tested before going live, and that is where the real problems were caught (see what broke). By the end it had 70 automated checks guarding it.

6Live with it, then tune it

The first real week showed what actually annoyed me. My morning brief moved from 8:00 to 9:30 because 8:00 was too early for how my days start. Design for how you really live, not how you think you live.

The rules that make it a secretary, not a spam machine

Tap each rule. These decisions matter more than the code.

What broke along the way

Every one of these looked fine in testing and failed in real life. I am naming them, not handing you the fixes. If you build this, you will meet them too.

1. The bot replied once, then went silent forever

Telegram kept re-sending my very first message and refused to deliver anything after it. The code was correct. The problem was how Google answers Telegram.

Look at the HTTP status code Google Apps Script returns, and what Telegram expects back. The fix lives outside Apps Script.

2. The newest AI model was the slowest

The first model I tried, the newest one on a free tier, took 90 seconds to answer, then refused with "high demand". Since I already had OpenAI API credits from my ChatGPT Pro plan, I used those instead. A secretary that takes 90 seconds is not a secretary.

Measure, do not assume. Time several models on the same real message, and never depend on just one.

3. Google Sheets quietly changed "15:00" into a date from 1899

My team's first booking request broke the Approve button. Nothing was wrong with the request. The spreadsheet had "helped".

Never trust a value to come back from a spreadsheet in the same form you wrote it.

The prompts

Two prompts, in the order it really happened: plan first, then build one small piece at a time.

Prompt 1: the planning prompt (the step most people skip)

Fill in the two lines in brackets and paste it into ChatGPT Codex. It does not write code. It writes a plan, ARCHITECTURE.md, covering what already exists, the riskiest assumption to test first, the stack, the data, the rules, the cost, a stage-by-stage build plan and the risks. Then it asks you the questions that would change the design. This is modelled on how my own assistant was planned.

I want to build a personal AI assistant for my work. Plan it with me before writing any code. About me: [2 to 3 lines: what you do, where people book you, which apps you already use] What goes wrong every week: [3 to 5 real problems, e.g. double bookings, forgotten follow-ups, my team asking me when I'm free] Work through these in order, then stop and wait for my approval: 1. Existing tools: check whether existing apps already solve this. For each one, say what it solves and exactly where it falls short for me. Only build what nothing else covers. 2. Riskiest assumption: name the one assumption that would break the whole plan if it is wrong, and give me a 20-minute test I can do by hand before any code is written. 3. Stack: recommend the simplest setup, using free tools and apps I already have, with no server I must maintain. For each choice, say why it beats the obvious alternative. 4. Data: what the assistant must remember, as tables with columns and one example row. 5. Rules: how it decides between a calendar event, a task, and a promise to someone; how reminders escalate; a daily message limit; quiet hours; what needs my confirmation; and Undo. 6. Cost: the monthly running cost at my usage, and what would make it start costing money. 7. Build plan: small stages, each testable on my phone before the next begins. Stage 1 only proves the connection works. 8. Risks: what could go wrong, stated plainly. Write it all as one document called ARCHITECTURE.md. Ask me any question whose answer would change the design, instead of guessing.

Prompt 2: the first build (the exact prompt from the reel)

Paste this into ChatGPT Codex. Once your plan is approved, this is a first stage to build. It builds version 1: every evening at 8:30 pm, tomorrow's calendar arrives on Telegram. It works on its own, costs nothing, and teaches you the setup every later stage builds on.

Write a Google Apps Script that sends me a Telegram message every evening at 8:30 pm India time, listing tomorrow's events from my Google Calendar with their start and end times. If tomorrow is empty, say "Nothing on your calendar tomorrow." Keep the bot token and my chat ID in Script Properties, never in the code. Include a function that sets up the daily trigger, and give me short setup steps in plain English at the end.
What version 1 does not do: it only talks to you. The two-way part, where you speak to it and it understands you, books your calendar, protects your Topmate slots and nags you about promises, is where the problems above live. That is the difference between a demo and something you rely on every day.

Want one built around your work?

Coaching institute, clinic, consultancy, tuition centre: the same pattern fits any work that runs on bookings, follow-ups and a small team. The design decisions are what make it fit your week, not the code.

Start tonightRun Prompt 1 to get your plan, then Prompt 2 for your first working piece.
Stuck on a step?Ask in the WhatsApp community below.
Build it properlyBook a 1:1 session and we will design yours.

Questions people ask

Do I need to know coding to build this?

No. ChatGPT Codex wrote most of the code. What you need is the ability to describe your problem clearly, test each step on your phone, and check what the AI built before trusting it.

Does it really cost Rs. 0 a month?

Yes for running it. Telegram, Google Apps Script, Google Calendar and Sheets are free, and the OpenAI API usage is covered by the free API credit that comes with ChatGPT Pro (the $100 and $200 plans include credit equal to the plan price), far above one person's usage. Building it used ChatGPT Pro with Codex, and that Pro access was provided through an agency campaign.

Why Telegram and not WhatsApp?

Telegram bots are free with no per-message charge and support buttons. A WhatsApp business bot needs a separate phone number and pre-approved message templates for reminders.

Is my data safe?

Everything lives in my own Google account. The bot answers only my Telegram account and approved team members, keys are stored outside the code, and it cannot send email at all. One honest trade-off: Messages go to an outside AI service, so I keep messages to names and times, never student reports or notes.

How long did it take?

About two days of building, in six small stages, after planning the design and testing the riskiest assumption first.

Can my team use it?

Yes. Team members can see when I am busy, never what I am doing, and send booking requests that I approve or decline with one tap.

Sources

Explore more

How students can adapt to AI Jobs safest from AI automation Student side hustles, real income Career counselling Study abroad guidance All resources
Connect With AptiGuide
📝Tally FormFree career queries 🗓️TopMateBook a 1:1 session 🔗LinkedInConnect with Anshul 🎬CareerTalks by AnshulDaily career reels and tips 📸AptiGuide InstituteInstitute updates and student results