Three games, built with agents.

All three run right here in the browser. Claude Opus 5.5 wrote the code. I wrote the prompts, chose what to focus on and reviewed the results. For GOLDENLINE I split the work into agent tasks; LITTLEBIG’s orchestrator did that itself.

01 · littlebigpr #49 · oct 2026
A tiny cartoon planet against deep blue space: a toy-bright downtown of towers and a clock tower rises on its curve, ringed by roads, red-roofed houses and green parkland, with puffy clouds and a striped hot-air balloon drifting past.
touch okLITTLEBIGA tiny cartoon planet with a capital and seven towns. Spin it, dive through the clouds and walk the streets, then ride any car, bus or ferry, see through a passer-by’s eyes, or take off as a bird. Everything is procedural: zero asset downloads./planet▶ play▶ play
LITTLEBIG — trailer · 0:30

Every shot is the game itself, rendered in the engine frame by frame.▶ play LITTLEBIG

Model
Claude Opus 5.5
Mode
1 orchestrator + 23 builders
Time
17 h, then v2 over 4 days
Code
~71k lines ts
Tests
49 spec files
Source
PR #49 ↗

How these were built.

01 · one-shot

Snake: one prompt, forty minutes.

A benchmark of sorts for a new model. Claude Opus 5.5 rebuilt the old 2D snake as a 3D game from a single prompt, natively, with no orchestration.

  1. prompt
  2. spec
  3. plan
  4. build
  5. critic ×3
  6. PR #43

A fresh critic agent scored 36 screenshots per round: 6 → 7 → 7.5 / 10

02 · orchestrated

GOLDENLINE: one orchestrator, ten agents.

I wrote the brief and defined the agent tasks. The orchestrator built the shared core and contracts, then ran the agents in two phases. Each agent owned one system and ran its own review loop.

briefmy prompt
orchestratorcore · contracts · review tooling · integration
phase a · parallel≤ 6 at once
A1atmosphere + post
A3beach & sand
A4pier
A5surface state
A6first-person player
A2oceanA7water shadingA8breaking waves
review gates · M2 beauty shot · M3 waves in motionlook pass
phase bsequential
B1surfing
B2polish · perf · startup
final reviewPR #44

03 · harnessed

LITTLEBIG: one prompt, a harness of agents.

One prompt from me, with the limits: quality first, small and fast, at most four agents at once. The orchestrator wrote the brief and split it into eleven builder tasks, then ran them in four phases. For v2 I asked for more life: other towns, roads, rides on anything that moves, bird flight and space. It wrote a second brief and ran twelve more builds.

Every build got a fresh critic that scored its screenshots at each altitude. Under 8.5 / 10 it went back for a refine round; all 8 world and life builds did.

critic → refine · every buildgates · art & wow · life & flow · perf & size

promptmine
orchestratorbrief · task split · integration · gates
phase 1 · foundation1 agent
F0engine · contracts · world gen · camera
phase 2 · world≤ 4 at once
A1terrain · ocean · nature
A2city
A3sky · clouds · light
A4camera
◆ gate G1art & wow · life & flow · perf & sizeowners fix
phase 3 · life and look≤ 4 at once
B1traffic
B2people
B3air
B4ink · post
◆ gate G2art & wow · life & flow · perf & sizeowners fix
phase 4 · siteparallel
C1site integration
C2load · size · perf
v2 · phase 1 · rides and region≤ 5 at once
R1region network
D1camera rides · bird
S1space · cloud fall
U1pop ui
L1followable life
v2 · phase 2 · towns and life≤ 2 at once
R2region rework
T1towns
H1roads
BFbird flight
BAbird animation
V1traffic · ferries
T2townsfolk
final reviewPR #49

My part

The agents wrote the code.
This is what I did, and for which game.

  1. 01Wrote the promptsall three
  2. 02Chose what to focus onall three · plus the limits for littlebig
  3. 03Defined the agent tasksgoldenline only · littlebig’s orchestrator split its own
  4. 04Reviewed the resultsall three
littlebig
BRIEF.md ↗TASKS.md ↗DECISIONS.md ↗PR #49 ↗
goldenline
BRIEF.md ↗TASKS.md ↗DECISIONS.md ↗PR #44 ↗
snake
spec ↗PR #43 ↗