I rebuilt a "dead" game's ranked scene as a live-service system — solo, 4 days, ~38k lines, 7 cron pipelines, a custom file-based database. GPT-5.5 and Claude Fable 5 built it as partners; the architecture was all mine
Post
I genuinely believe we're living the future, and this post is my evidence. Let me show you what I built, why, and who I am. (Apologies in advance for the caps and bolds — four days of 18-hour shifts will do that to a man.) Dirty Bomb is a free-to-play shooter game that officially stopped development years ago. The community never left. So I built it the ranked system it never got: dbranked.com . I'm a solo founder. No team, no contractors. Every system you're about to read existed in my head first — the whole architecture, the economy, the world, how the pieces connect. What I didn't have was a way to get something this complicated out of my imagination and onto a server. That's what the two AIs were for — and I mean two , as equals. GPT-5.5 and I started the build; two days in, Claude Fable 5 joined, and from there the three of us worked as partners — me as the architect, the two models as a build crew that never sleeps. The whole sprint was 4 days, 16–18 hours a day at the PC. I want to be honest about what that actually looked like, because "AI built my app" posts usually hide the interesting parts — starting with the fact that an AI cannot want anything. The wanting was all mine. What it is (this stopped being a website somewhere around day two — it's closer to a small live-service game that happens to run in a browser): An hourly ranking engine — dual-engine design (persistent skill + monthly seasons), dynamic rank distribution, placement pools, dormancy protection An integrity system — convicted cheaters get quarantined off the board, and every honest player who shared a lobby with one gets automatic "Hazard Pay" compensation, computed by the engine A daily contract system — players get personally-sized daily challenges, verified against their real match stats, paying a rank-rating surge and a permanent profile crown MERCSERV HQ — a walkable 2D "building" with a front-desk character, personnel files, floor chat, a daily-contract economy (RADS currency), and a cosmetics shop A Missions Hub — admin-authored missions players enlist in; the server snapshots baselines at signup, verifies progress hourly against actual gameplay, latches completions in-window, and settles payouts exactly once — including custom honor badges that land permanently on profiles An AI-personality news desk on the homepage — three merc characters running scripted episodes with a live ticker, feuds, audience interaction buttons, and a tic-tac-toe game you will lose Player intelligence layers — a sci-fi profile overlay and a full analytics page (activity heatmaps, form trends, rivals, smurf detection) A public patch log written like a studio changelog The stack will upset some of you: plain PHP, no framework, no build step, no npm, no SQL. Instead I have a custom file-based database the system built for itself: thousands of JSON files under a strict directory schema — per-player rating files, per-player hourly time-series histories , hourly full-ladder snapshots, season archives, a hazard ledger, a signal-state store, event/reward stores, user ledgers. Every write is atomic (tmp+rename), every payout is idempotent (write the result latch first, pay second — a crash mid-payout can never pay twice). Deployed by overwriting files in a shared host's browser file manager . No SSH. No git on the server. It runs. What's actually running in the background (this is the part nobody sees, and it's my favorite part): A live game-server scanner running every 60–180 seconds — watching real Dirty Bomb servers, recording who's playing where, and logging every session an honest player shares with a quarantined cheater (that's what funds Hazard Pay) The hourly core engine — pulls granular per-player stats from the community stats API, rebuilds the entire ladder, runs the season math, settles the daily contracts, settles missions, pre-generates player bios A 15-minute intel aggregator — distills the scanner's raw history into the live field intel you see on the site, touching zero external APIs A 5-minute desk regenerator — keeps the homepage talk show on air, plus a separate pipeline that banks evergreen AI banter so the show never runs dry An anomaly + smurf watcher — reads per-snapshot rating-breakdown margins across the whole history looking for veteran patterns on fresh accounts A statistics engine for the analytics deck — form trends, forecasts, comparisons, all computed in cron so pages stay instant; heavy math never runs on page load That's 7 independent cron pipelines feeding one site, coordinating through nothing but JSON files and discipline. The core engine library alone is ~6,000 lines and 210 functions; the player-facing API speaks 18 actions, every authenticated one behind HMAC-signed tokens and per-user throttles. How the two AIs actually divided the work: First, the part most posts get backwards: the architecture didn't come from the models. The dual ranking engine, the quarantine + Hazard Pay idea, the in-universe HR-department world, the missions lifecycle, the badge honors — all of that was designed in my head, often years of knowing this community condensed into "here is exactly what I want and why." The models' job was to take a system that existed only as my thoughts and make it real : structured on paper first, then in code. That translation step — imagination → formal architecture → running system — is what was genuinely impossible for one person before, and it's what made me write this post. GPT-5.5 was there from hour one and did serious heavy lifting: the first two days of the sprint were me and GPT turning the vision into foundations, structure, and working systems. This was not a "brainstorming assistant" — it built. When Claude Fable 5 joined on day three, I didn't switch models — I gained a second builder , and the two worked as genuine partners on my project: one would design or implement, I'd carry the result to the other for review, and back. Two frontier models examining the same system from different angles is the most powerful QA setup I've ever had. When they agreed, I shipped. When they disagreed, I'd just found the exact spot that needed my judgment. What Claude (Fable 5) added on top of building was verification as a discipline : it would spin up a local copy of the site, generate synthetic test data, render pages in a headless browser, screenshot them, click through flows, run the cron, and check the payouts landed in the right ledgers before handing me files. It caught its own bugs this way (a payout that silently didn't fire in cron context, a settle that would have closed missions with zero winners if a dependency failed to load), and diffed every changed file against production to prove nothing else was touched. I'm not going to rank the two models against each other, because that would be lying: this system exists because of both. Different strengths, same standard. What I actually did , since "I did it alone" needs defining: the entire vision and architecture, every product decision, the world-building and tone (the whole thing is written in-universe, like a mercenary HR department), all the QA on the real site, every deploy, and the taste calls — what's fun, what's fair, what's honest. The hard rule I enforced on both models: every public claim on the site must be literally true. No fake stats, no invented testimonials, no "AI slop" copy. The models are extraordinary executors — but an executor needs something to execute. That was the four days: me emptying my head, them keeping up. Honest limitations: it took iteration. Models confidently produce code that lints clean and is still wrong in production context — the only cure was making verification non-negotiable. And neither model replaces knowing what you want. Site: dbranked.com — the patch log on it documents everything above. Happy to answer anything about the workflow, the prompting, the architecture, or why a sane person builds a game economy on flat files. (Screenshots in comments / attached.)
Intent Score
0
Intent
99
Confidence
Summary
The post is about building a game ranking system with AI and has no connection to windows or home repair.
Reasoning
This is clearly a software/product/AI build post about a game service, not a homeowner discussing window problems, replacement, or comparison shopping.
Extracted Signals
- software project
“I rebuilt a "dead" game's ranked scene as a live-service system”
- AI-built app
“GPT-5.5 and Claude Fable 5 built it as partners; the architecture was all mine”
- game service
“Dirty Bomb is a free-to-play shooter game that officially stopped development years ago.”
Model: gpt-5.4-mini · Prompt: v3 · 6/12/2026, 9:01:56 AM