(0:00) [The 21-second clip starts by itself; let it run, then pause 3 seconds.] Good morning. You might be wondering why I am showing you an Excel sheet. Well, it is not Excel. Three days ago this app did not exist. I did not type a line of its code: a coding agent built it from a conversation. I chose what it should do and checked what came back. By the end of this talk, I want you to have one project you actually want to start.
RO AI Labs

I Love Tech Bucharest · 31 October 2026

How Agentic AI Coding
Democratizes
Application Development

ship-it-with.ai

Mihai Cvasnievschi · Founder, RO AI Labs

Cover of the manual Ship It With AI
(1:00) I’m Mihai. I wrote the free manual ship-it-with.ai, which accompanies this talk. [Pause for 2 seconds.]
Mihai Cvasnievschi

Who is talking

Mihai Cvasnievschi

25+

years of professional
software development

10+

years building
AI systems

Founder of RO AI Labs

(1:10) I’ve built software professionally for over twenty-five years, and AI systems for over ten. I’ve used computers for the past forty-two years and sold my first program at sixteen. I run RO AI Labs in Bucharest. Today I’ll show you how I build with AI and check the results. [Pause for 2 seconds.]

The questions everybody asks

1

Can I really
write code with AI?

2

How do I
start?

3

Which AI agent
do I choose?

4

Is there a place for me
in the age of AI?

(1:35) These are the four questions I am often asked. Show of hands: who has written code for work, study or a personal project? [Allow 8 seconds.] Now, who has tried coding with an AI agent? [Allow 8 seconds; observe without predicting the result.] Whether you already write code or are learning, these questions apply. Can I really write code with AI? How do I start? Which AI agent do I choose? And is there a place for me in the age of AI? We’ll look at briefing, reviewing and testing an agent’s work. First, why do these questions matter now? [Pause for 3 seconds.]

Why these questions matter now

Three numbers from 2026

US job ads

73%

of US tech job postings
ask for AI skills.

Up from 15% in January 2024.

Advertised pay · global

+62%

higher advertised pay for
jobs that ask for AI skills.

Job ads compared, not people.

US workers aged 22–25

−19%

employment in jobs
most exposed to AI.

Compared with keeping pace
with less-exposed jobs.
A pattern, not a proven cause.

Sources: Dice, Jul 2026 · PwC AI Jobs Barometer, Jun 2026 · Stanford Digital Economy Lab, Aug 2026
www.dice.com/hiring/recruitment/reports/tech-hiring-myths-vs-reality · www.pwc.com/gx/en/1/services/ai/ai-jobs-barometer.html · digitaleconomy.stanford.edu/news/canariesaug26/

(2:40) Three numbers from this year. First, what employers ask for: Dice reports that seventy-three percent of US tech job postings asked for at least one AI skill in May, up from fifteen percent in January 2024. “AI skills” includes much more than using a coding agent, and these are job ads, not people hired. [Pause 2 seconds.] Second, advertised pay: PwC’s global study found sixty-two percent higher advertised pay for jobs asking for AI skills. This compares groups of job ads, not the same person before and after learning AI. [Pause 2 seconds.] Third, younger workers in the US, aged twenty-two to twenty-five: in highly AI-exposed occupations, employment was nineteen percent below where it would be if it had kept pace with less-exposed jobs. [Pause 3 seconds.] If you are a junior developer, the skills you are building now also help you check an agent’s work: reading code, testing changes, understanding the problem. Use them from day one. [Pause 3 seconds.]

It’s Halloween, so here is a horror story

$10,000
a week

Senior devs offer to clean up
your vibe-coded app.

The cleanup offer

3 senior engineers

1 week

Cut the code agents duplicated

Keep the app working

Their example: 14 date formatters
that should have been one

Source: Slopfix: three senior engineers, Hacker News front page, July 2026
odra.dev/slopfix

(4:10) There is a second risk: using AI without steering it. It’s Halloween, so here is a horror story: an app that works, until you try to change it. Slopfix, three senior engineers, advertises a week of cleanup for ten thousand dollars. Their example is fourteen date formatters where one would do. [Point to the diagram; 4 seconds.] Each copy is another place to fix the same bug. Their offer is to remove duplication and keep the app working. They start by writing down what the app does, then check that it still does it after the cleanup. [Pause 3 seconds.]

The twist, in their own words

“Yes, we use Claude Code too.
On a very short leash.”

Same AI. More experience steering it.
You can learn to steer it too.

Source: Slopfix
odra.dev/slopfix

(5:00) Now the twist, from their own website: “Yes, we use Claude Code too. On a very short leash.” [Pause 3 seconds.] The people offering that cleanup also use Claude Code. They credit thirty years of combined engineering experience, and they apply it by writing down what the app should do, setting boundaries and checking the result. That is steering, and you can learn it. [Pause 3 seconds.] Start by understanding what work you can give an agent, and what you will need to check.

What changed in five years

How much the AI does

Autocomplete

a line

Chat

a function

Agent

a change across files, with tests

2 in 3 surveyed developers
use a coding assistant or agent.

Source: Stack Overflow Developer Survey 2026 · AI-use question: 17,464 responses; 65.9% use a coding assistant or agent
survey.stackoverflow.co/2026/ai

(5:55) What changes is the size of the task you can delegate. Autocomplete suggests the next line. Chat can explain a function or write one for you. An agent can read files across a project, make changes, run the program and test the result. You give it a goal and boundaries. Two thirds of respondents to Stack Overflow’s AI-use question already use a coding assistant or agent at work. That includes autocomplete. If you haven’t tried one yet, everything you already know about code still counts. Start with one change you know how to check. [Pause 4 seconds.]

Linus Torvalds, creator of Linux, on AI · 13 October 2026

Don’t shy away
from AI.

From: Linus Torvalds · To: Laurent Pinchart

Subject: Re: [PATCH] docs: remove Assisted-by references

“The people I respect and call intelligent are not the ones
that shy away from using new and interesting tools.”

“I expect people to understand what they are doing.”

Source: Linux kernel mailing list, 13 Oct 2026 · debate over requiring human-written commit messages
lore.kernel.org/all/CAHk-=whTs09_ozN+e09GU68N1wPP56MUzDjrg9rnKxQUa0A_4Q@mail.gmail.com/

(6:45) In an email earlier this month, Linus Torvalds, the creator of Linux, responded to a proposal to require human-written explanations of code changes. His point was that using AI does not mean you lack understanding. [Pause 2 seconds.] He welcomes new ways of working, and he expects people to understand what they are doing. Whatever help you use, understand what you ship. [Pause 3 seconds.]

ship-it-with.ai, in three lines

The agents write the code.
You understand the problem.
You remain accountable
for the judgment.

The opening lines of the manual.

QR code linking to ship-it-with.ai

ship-it-with.ai

Free, in English and Romanian

(7:30) Take a moment to read this. [Silence: 10 seconds.] The agents write the code. You understand the problem. You remain accountable for the judgment. A nurse knows which paperwork wastes time. A logistics planner knows which Excel file everyone dreads. An engineer knows which parts of an existing app must keep working. Start with a problem you understand. Use what you know; learn what you need to check the result. [Allow 10 seconds for photos and scanning.] The code takes you to the free manual. Here is how I did it with the Excel clone from the opening.

Question 1 · Can I really write code with AI?

Wednesday evening · the first message

› I want a browser version of Excel.
  Before building anything, propose
  the prompt I should give you.

How long until it works?

10 min

1 hour

1 day

(8:30) Question one: can I really write code with AI? This is the first message I sent for the Excel clone: propose the prompt before building. Read the brief: what did it assume, what did it leave out, and what would count as done? Now your prediction: ten minutes to a working prototype? [Hands: 5 seconds.] An hour? [5 seconds.] A day? [5 seconds.] Remember your answer. Watch the real clock in the corner. [Pause 2 seconds.]
(9:25) [The 2:03 video starts by itself; the cues below are video time.] [0:00–0:04, title card; speak immediately.] Codex with Superpowers. [0:04–0:12, first human message: stay quiet.] [0:14, brief.] The brief includes worksheets, formulas and autosave. [0:28, brief review.] I keep the scope: no server, and tests plus a correct calculation define done. [0:36, design review.] A second reviewer checks the design. [0:41, the spec.] The spec says what to build. I read it before approving. [0:49–0:55, approval: stay quiet.] [0:57, the plan.] Review the plan before it runs. [1:04, execution.] Tests first, then code. Watch it work; stop it if it drifts. [1:14, review.] Read what it found: two bugs. Each gets a failing test first. [1:21, verification.] Thirty-eight tests pass. Check the proof. [1:30–1:38, final human reply: stay quiet.] [1:41, browser result.] Try it yourself: SUM gives sixty, then sixty-five after an input changes. [On the end card at about 1:58.] Forty-five minutes. Four human messages. [Let the video finish.] Whoever said an hour, you were closest. Excel is forty-one years old. This one is three days old. A prototype, written by an agent. [Pause 4 seconds.]

What you just watched

At every step, you do something

Your prompt starts it. Then, for each step:

1

Research

Agent: Writes the spec

You: Read the spec

2

Plan

Agent: Writes the plan

You: Review the plan

3

Execute

Agent: Writes tests and code

You: Watch. Stop it
if it drifts.

4

Review

Agent: Reviews the code

You: Read the
findings

5

Verify

Agent: Runs the checks

You: Check the proof
yourself

6

Ship

Agent: Prepares the commit

You: Decide whether
to ship

ship-it-with.ai · Chapter 4, The six-phase loop

(12:02) Your prompt starts the loop: research, plan, execute, review, verify, ship. [Point along the six cards; 4 seconds.] The agent writes a spec: read it before you approve. It writes a plan: review it before it runs. Watch the work and stop it if it drifts. Read what the review found. Check the proof and try the app yourself. The agent prepares the commit; you decide whether to ship. Four messages in the recording, and still something to do at every step. [Pause 4 seconds.]

Question 1 · Can I really write code with AI?

Four skills for steering AI

Brief

Say precisely
what you want

Read

Read what
comes back

Test

Test before
you trust

Judge

Know when to
say “not yet”

ship-it-with.ai · Chapters 3, 5 and 8

(12:52) Can I really write code with AI? Yes. If you already read code and test changes, bring those skills to the agent’s work. If you are learning, practise them on a small task. Brief: explain the problem, who it is for, and what done looks like. “Make me an Excel clone” is a starting point. “Keep it in this browser, save my changes, and make this calculation work” gives you something to check. Read: inspect what comes back, including the choices the agent made. If it says “saved locally”, ask whether that means it survives a refresh or appears on another computer. [Pause 3 seconds.] Test: write the expected behaviour before building. Use tests to record that agreement, and read them as well as the code. Try examples whose answers you know, including an empty input. Judge: decide whether the result is ready, the task needs to be smaller, or you need another person’s review. [Pause 3 seconds.] Brief, read, test, judge. Apply them to one small change. [Pause 3 seconds.]

Question 2 · How do I start?

Your first three messages

Your agent · a first project

1

Your problem,
in your words

› Our team tracks shared expenses in an Excel file.
  Help me plan a small app. Ask me questions first.

2

How you will
check it

› Before any code, list examples I can try,
  each with the answer I should see.

3

The smallest
version

› Build only adding expenses and their total.
  I will try it before we add anything else.

Practice example · made-up data · ship-it-with.ai

(14:22) Question two: how do you start? With three messages. [Allow 5 seconds to read.] First, your problem in your own words: our team tracks shared expenses in an Excel file; help me plan a small app for it, and ask me questions first. I started the Excel clone the same way: before building anything, I asked the agent to propose the brief. [Pause 2 seconds.] Second, how you will check it: before any code, examples you can try, each with the answer you expect. Ten lei plus twenty lei should show thirty. Ask the agent to turn those examples into tests. [Pause 2 seconds.] Third, the smallest version: only adding expenses and their total. Try it yourself, then add the next thing. For an existing project, also say what must keep working. Use made-up data, and keep passwords and connected accounts out of the agent’s reach. [Pause 3 seconds.]

Question 3 · Which AI agent do I choose?

One that fits your project
and lets you check its work.

Things that keep changing

Model names

Interfaces

Default settings

Usage limits

Did not change

Control what it can see

Control what it can change

Control how you check it

Control where you apply it

Source: OpenAI Codex changelog · checked 30 Oct 2026
learn.chatgpt.com/docs/changelog

(15:42) Question three: which agent? One that fits your project and lets you check its work. Here is a confession. I started writing the manual at the beginning of this year, and some of it is already old. Model names, interfaces, default settings, usage limits: these keep moving. What stays useful is the right-hand column. Context: what can it see? Actions: what can it change? Verification: how will you check the result? Application: where does it make sense to use it? [Pause 4 seconds.] Learn those questions and you can take them into the next agent. Check its access and permission settings before you start. [Pause 3 seconds.]

Question 4 · Is there a place for me in the age of AI?

The agent assumes. You know.

If you build

Unless you say otherwise, it may assume

In Romania

A shop

Prices like $1,234.56

1.234,56 lei

Invoicing

Standard VAT 19%

21% since Aug 2025

An HR app

Easter on 28 Mar 2027

Orthodox Easter, 2 May 2027

A booking app

Dates like 10/31/2026

31.10.2026

A delivery form

ZIP code, 5 digits

Postal code, 6 digits

The agent proposes the flows. The details are your job.

Sources: Fiscal Code art. 291 (ANAF) · Labour Code art. 139 · Unicode CLDR · Universal Postal Union · Dollar default: Miao et al., arXiv, June 2026
arxiv.org/abs/2606.17688v1

(16:42) Question four: is there a place for me in the age of AI? Laurie Voss makes this point: code alone does not describe how a workplace runs. An agent can fill missing details with defaults. A June 2026 study showed models a photo, like a chair, with no clue where the user lives, and asked the price: one model answered in US dollars ninety-eight percent of the time, and the others were similar. Five things you might build. A shop: here it is lei, with a decimal comma. Invoicing: our standard VAT rate has been twenty-one percent since August 2025; older sources say nineteen. An HR app: the Easter holiday follows Orthodox Easter, the second of May in 2027, and employees of other Christian churches get their own Easter dates. A booking app: dates here are written day first, thirty-one, ten, twenty twenty-six: today. And a delivery form: our postal codes have six digits, not five. [Pause 3 seconds.] [Optional, for founders:] If you are thinking of a startup, this is your advantage: what you know about your market and your customers is what the agent does not have until you write it down. Spelling out those details is your job: in the brief, checked with the people who do the work, and checked in the result. [Pause 2 seconds.] So, what can you do tomorrow?

Before the weekend is over

Start tomorrow

One agent

Any coding agent.
You set its limits.

One project

A small app or fix.
Made-up data.

One rule

Never ship what
you cannot explain.

The manual is free: ship-it-with.ai

Manual in English and Romanian

(18:25) Tomorrow: one agent, one project, one rule. Any coding agent will do: the one you already use, or one that fits your project. Before you start, check what it can see and what it can change. Choose one small app or fix, with made-up data, like the expense app. Write the brief and the checks before building. Ask someone to try it without your help, and repeat the checks after every fix. [Pause 3 seconds.] Brief, read, test, judge. Never ship what you cannot explain. The manual is free. [Pause 4 seconds.]

The four questions, answered

1 · Can I really write code with AI?

Yes. Start with
four skills.

2 · How do I start?

One small change,
one evening at a time.

3 · Which AI agent do I choose?

One that fits
your project.

4 · Is there a place for me in the age of AI?

_

(19:15) I promised four answers. Can I really write code with AI? Yes: brief, read, test, judge. How do I start? One small change, one evening at a time. Which AI agent? One that fits your project and lets you check the result. [Pause 3 seconds.] Pick a task and write the checks. Is there a place for you in the age of AI? [Pause 3 seconds; advance.]

Is there a place for you in the age of AI?

Build one.

ship-it-with.ai

(19:50) [Hold the slide silently for 3 seconds.] Build one. Or improve one. Tomorrow, make one small change with an agent, and check what comes back. Thank you. [Allow 6 seconds for applause before advancing.]
RO AI Labs

Questions

Thank you

Let’s talk more at Let’s Talk AI #6

Wednesday 4 November, 18:30

Zăganu, Calea Victoriei 91-93, București

Free, in person · sign up on Luma or Meetup

lets-talk-ai.ro · ship-it-with.ai

QR code linking to Mihai Cvasnievschi’s LinkedIn profile

Connect on LinkedIn

linkedin.com/in/mihaicvasnievschi

(20:13) Thank you. We have time for questions. [Take questions as the remaining time allows, following the host’s cue.] If we don’t get to yours, let’s talk more on Wednesday at Let’s Talk AI, the monthly AI mixer I organize here in Bucharest: the fourth of November, from half past six in the evening, at Zăganu on Calea Victoriei. It is free; sign up at lets-talk-ai.ro. To stay in touch, scan the code to connect on LinkedIn. The manual is free too, in English and Romanian. Thank you. [Leave this slide up while people leave.]