War Trader title screen: a trader on rainy Wall Street above a sewer full of rat-men at computer terminals
Work archive06 · Game / AI-directed production

Version 3 · Godot 4.7 · Windows build

War Trader战神交易员

A pixel-art action trading game set in a 1980s New York. Trade a live market by day, fight rat-men in the sewers by night, and turn stolen intel into tomorrow’s edge.

Role
Creator · Game and economy design · AI-directed engineering · QA
Format
V3 playable end to end · Windows build

Design question

Can trading and fighting feed each other?

Each of the seven days is a loop: read the morning brief, trade ten fictional stocks, spend profits on gear and skills, then go underground to steal the intel that moves tomorrow’s market. About 28% of that intel is forged.

I directed the build with AI coding agents—49 studio-role agents and 119 installed skills—while I owned the design, the rules, and every acceptance test.

15,183Tests passing, 0 failures
16.9KLines of GDScript
173Music tracks and sound effects synthesized in code
19Endings, tracked in an in-game gallery

Highlights

A market with causes, a city with consequences.

01

Prices move for reasons.

Ten fictional stocks react to scheduled events and four multi-day event chains. A “why it moved” panel breaks every price change into the forces pushing and dragging it.

02

Intel you have to doubt.

Clues state facts, never direction, and about 28% are fake. A seven-day timeline lets you link, flag, and cross-check them.

03

Nights that change the next day.

Five six-zone routes with ladders, switches, branches, and hazards. Pistol, shotgun, a skill tree, three rat factions, two bosses—and choices underground change tomorrow’s market.

04

Tested like a product.

Balanced by simulation: 2,000 market games per difficulty and bot fights from Easy (8/8 won) to Nightmare (2/8). The exported build stutters past 100 ms only 3 times in a full seven-day run.

How I used AI

One person, an AI studio.

  1. Chose the team

    Picked five open-source skill libraries (each 500+ GitHub stars) and installed 119 skills and 49 studio-role agents scoped to this project only.

  2. Ran it test-first

    Plan → failing test → fix → verify. Every bug the bot or layout checker found got a regression test before it was fixed.

  3. Kept one source of truth

    Rules, money limits, difficulty, and combat values live in written docs and JSON, so every agent session started from the same facts.

Technical foundation

Deterministic underneath, loud on top.

Engine
Godot 4.7.2 · GDScript · 640×360 pixel art, integer scaling
Systems
Event-driven market · seeded RNG · integer-cent accounting · versioned saves
Audio
39 chiptune tracks + 134 sound effects synthesized by a Python script
QA
Headless test suite · per-screen layout checker · seven-day bot smoke test · balance sweeps

Version 3

Playable end to end, not yet published.

V3 adds ten stocks, a fixed story order with four event chains, skills, visible weapons, and five six-zone routes. Honest limits: the routes still use the sewer tileset tinted per route (the V3 map art appears in zone banners), ladders reuse jump frames, and music and sound effects are synthesized placeholders until final audio is commissioned.

Back to all workRead the project logDiscuss the project