We’ve all heard the stories about AI and how much faster it supposedly makes you at building software. Some people say they’ve become 10x or even 100x developers. Others say LLMs produce garbage, code nobody can maintain, and a steady stream of slop.

While I’ve read plenty of those arguments, what I’ve almost never seen is a number. So I ran the test on myself.

There’s an app I’d been meaning to build as a side project anyway. And to be completely transparent, part of this post is me marketing it. A small part. The numbers are why I wrote it.

What’s being built

It’s a mobile app where users can play multiplayer mini games with some twist. One of the games is Battleships with additional power-ups (bombs, mines, torpedoes). Users can play for free or they can make things a bit interesting and stake crypto. The winner wins the staked amount and the platform takes a small fee. The fees are spent on app maintenance, prize pool for tournaments/rankings and for referrals.

The name of the app is Kombat, currently its in closed testing phase. If you want to give it a try, head over to playkombat.xyz and apply, we’ll send the invitation as soon as possible. Currently its only IOS supported with Android version coming soon.

On the surface, this look really simple. The game logic can be explained by a small child. But thats not really the case, the app itself is pretty complex for 1 man team.

It consists of:

• Mobile app with 17 different screens and 5 games. Non custodial crypto wallet. WebSocket connection management and all the different states that can happen between games, disconnects, reconnects, timeouts etc. Notifications, user profile, stake pots, settings, multiplayer lobby, free / staked games etc.

• Backend server that exposes both REST and WebSocket. In order for users to not cheat, the whole game logic is handled on the server side. From joining games, gameplay, timeouts, disconnects, left rooms and a plethora of different edge cases. It also handles authentication, users, statistics, and the staking that manages funds.

• Webapp that has a landing page and a feature that allows users to stake directly from the browser with their own wallet, plus an admin panel that allows looking into statistics and management of a server.

• Designs, from game screens and models to custom icons and the landing page.

Where I’m starting from

I’ve been a software engineer for about eight years. Mostly backend and infrastructure. I’m comfortable with backend servers, databases, queues, clouds, Kubernetes, Docker and the general business of keeping a server up.

I’ve done frontend when I had to. I don’t enjoy it and I don’t think I’m particularly good at it. I’m definitely not a designer. Also, I’d never built a mobile app. Zero React Native (or any other mobile framework) before July 1st. I picked React Native because I already had ReactJs experience and knew how things worked.

The measured part

Everything below comes from git log on the repo.

| First commit | 2026–07–01, 00:22 |
| Last commit. | 2026–08–04, 18:20 |
| Elapsed | 34 days |
| Commits | 296 |
| Production code | 58,696 lines (TS, TSX, Solidity, CSS, SQL) |
| Total churn | 127,511 added, 23,038 deleted |
| Contributors | 1 |

For now, I will be keeping the repository private, but if someone is interested, I am willing to share it on an individual basis.

Here’s every commit bucketed by hour of day:

00:00 ███████████████████████████████████████████ 43
01:00 ██████████████████████████████████████████ 42
02:00 █████████ 9
03:00 to 10:00 nothing
11:00 ███ 3
12:00 0
13:00 ██ 2
14:00 █████████ 9
15:00 ███████████ 11
16:00 █████████████████ 17
17:00 ██████████████████████████████ 30
18:00 █████████ 9
19:00 ██████████████ 14
20:00 ██████████████ 14
21:00 ███████ 7
22:00 ██████████████████████████████████ 34
23:00 ████████████████████████████████████████████████████ 52

86% of commits land between 17:00 and 02:00, after my work day. The rest are weekend afternoon commits.

To turn that into hours I clustered commits into sessions, treating any gap over four hours as a new session. That gives 34 sessions and 70.4 hours of commit-to-commit span. Commit spans undercount, since they capture nothing before the first commit of a night or after the last, so I added half an hour per session for setup and wind-down.

Total: roughly 85 hours. Median session about 1.5 hours. The longest was 8.3 hours on July 31st, 52 commits, straight through midnight. That total is commit sessions reconstructed from git history. It doesn’t count the thinking, reading and planning that happened away from the keyboard.

Seven days I didn’t touch it at all.

85 hours is a bit over two full-time working weeks, spread across a month of evenings. That’s also a useful thing to know about myself independent of this whole exercise: my real capacity for a side project is about 80 hours a month, and no amount of enthusiasm changes that.

What this would have cost me without AI

Now the estimates.

The assumption is the same developer, me, with the same knowledge, writing everything by hand with no AI assistance. Design is excluded here and priced separately, since I couldn’t have produced it at any time cost.

As any dev knows, estimates are tricky, and they get worse when you’re not the expert on the thing you’re building. I have no mobile experience and no game experience, so my first pass at these numbers was almost certainly too high. So I will first do one estimation and cut time by 20%.

| Area | | Full-time weeks |
| Mobile shell: nav, auth, wallet, onboarding, etc | 2.5-3 weeks |
| 5 animated game boards in Skia/Reanimated | 5-6 weeks |
| Socket sync, reconnect, match resume, errors,retries | 1 week |
| Server: auth, JWT, users, wallets | 1.5 weeks |
| Server: socket gateway, room lifecycle, presence | 1.5 weeks |
| Server: 5 game integrations, sweeper, forfeits, etc. | 2.5 weeks |
| Staking and chain service | 2 weeks |
| Admin, emails, maintenance, cache, migrations | 1 week |
| Shared game rules, board topology, bots | 2.5 weeks |
| Solidity: contracts, Foundry tests, deploy scripts | 2-2.5 weeks |
| Web app, with admin panel | 1 week |
| CI/CD, Docker, registry, deploy pipeline | 1 week |
| Specs, planning, docs | 1 week |
| Total | ~25 weeks |

25 weeks at 40 hours is about 1,000 engineering hours. Minus the 20%, we have 800 hours.

Become a Medium member
Since I do not have time to work on the app full time and we’ve seen previously that I have up to 80 hours to spend per month on part time project:

800 hours ÷ 80 hours per month = 10 months

Add the re-context tax. Coming back to a Redis lock or a socket reconnect flow cold after four days off costs re-orientation time, and on side projects that’s conservatively 10 to 20 percent.

I’d estimate this to be 11 to 12 months project.

                           | With AI | Without AI |

| My own hours | ~80 | ~800 |
| In months | 1 month | ~11-12 months |
| Cash spent | $100 | $15k-50k+, for designs

The design problem, which is a money problem

I want to separate this out because it’s the part I genuinely could not have solved with more evenings.

I can’t design. If AI hadn’t done it, the money was leaving my account. There is no version of this where I sit down for another six months and produce a brand identity.

| Line item | Freelance | Studio |

| Brand identity plus full icon and asset kit | $2-3k | $20-50k |
| Mobile UI: 17 screens plus a design system. | $5-10k | $20-40k |
| 5 game board designs (with all screens/states) | $5-15k | $25-60k |
| Web app design | $1-3k | $8-15k |
| Total | $13-31k | $70-165k |

I don’t have much experience with design pricing, but I’ve asked designers for quotes before and never got anything below $1–3k for a couple of screens and modals. That’s for good ones, not Fiverr.

And if I’d just hired someone to build the whole thing

Roughly 6.5 person-months of engineering (mobile around 2.5 months, backend 2, Solidity 0.5, DevOps 0.5, plus project management overhead), which a small team would run in 2 to 3 calendar months. These are rate-based estimates using standard market rates, not a quote I actually received.

| Tier | Eng. rate | Engineering | Design | Total |

| EU / remote blend | $60–100/hr | $60–100k | $25–45k | $85–145k |
| US agency | $150–250/hr | $150–250k | $70–165k | $220–415k |

Solidity bills above blended rate, so add $10–20k to either row.

So thats 85,000–400,000$ versus $100 Claude subscription plus a month of my own time.

If I were to hire an agency or a freelancer, I would have also had to spend my time, doing testing, reviews etc… Probably not as much time as I’ve spent writing everything myself but close.

What AI did not do

If I only wrote the good parts this would be an advertisement.

The biggest thing AI did not do is the part people assume it did: decide. This was not vibe coding. Work went in piece by piece. I decided how each part should work, wrote it down, and reviewed/tested what came back. To be transparent, I have not read every single line of code, but I did review 100% of code thats most important, game and websocket connection logic, staking etc.

AI also did not come up with tests, he just wrote them. I told it which ones to write and what they had to assert, because a suite the model invents for itself mostly proves the code does what the code does. The model typed most of the code. It made almost none of the decisions.

It didn’t run the infrastructure either. Servers, Database and Redis, DNS, secrets, TestFlight builds, deploys: all me.

What actually worked

Five things, roughly in order of how much they mattered.

A real test suite, written early, that the model has to keep green. Without it you get a model that fixes your bug and quietly breaks two things three files away, and you find out four days later. With end-to-end coverage on room lifecycle, game flow, timeouts and session expiry, the feedback loop closes in seconds instead of days.

A bug log the model maintains itself. I have three files in the repo: SERVER.BAI.md, MOBILE.BAI.md, WEB.BAI.md (BAI = bugs and inconsistencies/improvements). The instruction in my project config is: if you notice a bug, dead code, an inconsistency or a potential improvement that is not part of the task you were given, do not silently fix it and do not ignore it. Log it there with the date, the file and line, why it’s wrong, and a suggested fix. Then tell me.

This solved a problem I didn’t anticipate. Left alone, the model wanders. You ask for one thing and it “helpfully” refactors four unrelated files, and now your diff is unreviewable. Giving it a place to put those observations means it stays on task and I get a queue of real findings to triage on my own schedule. Some of them were things I’d have missed entirely.

I decided the architecture. Always. Cache layout, lock granularity, whether stakes settle on-chain or in a ledger, the socket protocol shape. I made those calls, wrote them down, and then handed over implementation with step-by-step detail. When I got lazy and gave a vague instruction, I got a vague result and paid for it in review time.

Specs as files, not chat messages. For anything non-trivial I wrote a markdown spec into the repo first, argued with it, then implemented against it. There are 63 markdown files in the repo. They’re also the reason I could pick the project back up after four days off without losing the thread.

Redesigning things from scratch, on purpose. Staking went through a full rewrite from per-match escrow contracts to a pooled pot. The mobile UI got rebuilt around a proper design system with an enforced spacing scale after the first version drifted. Both were multi-week decisions in the old world. Here they were one evening each, which meant I actually made them instead of living with the first design forever.

So whats the actual ROI of 100$ Claude Code subscription?

As we’ve seen, I’ve spent 85 hours vs 800 hours. I’ve also spent 100$ versus 15,000$ + that I would have had to spend on designs. So in my case, its easy to calculate ROI, its almost a year of my time and 15,000+ dollars

A 12-month side project doesn’t ship late. It doesn’t ship at all. Expo SDKs move, React Native’s architecture shifts, dependencies rot, and somewhere around month nine (but realistically much sooner) you stop opening the folder. I have a graveyard of those. Most developers do.

If I calculate the numbers and just take hours into consideration, this month Claude Code made me 10x developer, under ideal conditions: one person, a brand-new codebase, an app in test mode with no real players. Hand me an inherited app with live players, or a big enterprise codebase, and the multiplier drops hard. This post measures the best case.

If you find the app interesting and want to test it, suggest improvements or new games, or contribute in any other way, Visit playkombat.xyz, apply to receive an invitation or join our Discord or Telegram group.