BuilderBuilder
Open source · MIT · v0.1

Production skills for AI-built games.

Reusable Claude Code workflows for the failures that green tests and pretty renders miss: contradictory references, weak geometry, effort dilution, broken gameplay and self-certified work.

Install the skillsView on GitHub
Skills
7
Production gates
5
Cross-model subagents
2
Licence
MIT
Version
v0.1
Installation · RemakeBench Skills v0.1

Install once. Keep the rules close.

The managed plugin is the shortest route. Manual installation keeps editable copies in your Claude Code skills directory.

Claude Code · terminal
claude plugin marketplace add RemakeBench/skills
claude plugin install remakebench-skills@remakebench
Claude Code · session
/plugin marketplace add RemakeBench/skills
/plugin install remakebench-skills@remakebench
/reload-plugins
Manual installation
git clone https://github.com/RemakeBench/skills.git
mkdir -p ~/.claude/skills
cp -R skills/skills/* ~/.claude/skills/
What runs where
  • The five production skills need no external model CLI.
  • codex-subagent and grok-subagent are optional.
  • Their wrappers require jq and the corresponding authenticated CLI.
  • External model calls may use account quota or incur charges.
Production flow

Proof enters before polish.

Each stage hands evidence to the next. A playable build arrives early; independent acceptance remains active until shipping.

  1. 01Reference pack
  2. 02Congruence gate
  3. 03Playable greybox
  4. 04Asset build + independent judge loop
  5. 05Assembly
  6. 06Real-input playtest
  7. 07Ship evidence
Public catalogue

Seven skills. Two jobs.

v0.1 turns seven recurring AI game-production failures into reusable gates. It is deliberately small, opinionated and expected to improve as more builds expose new failure modes.

01

Production gates

5 skills
reference-pack-authorityProduction gates

Stop contradictory concept art from becoming contradictory geometry.

Generated views can disagree on counts, layout, scale and elevation. This gate creates one written spatial authority before modelling begins.

View skill on GitHub
game-production-stagesProduction gates

Build in the right order, with an evidence-backed exit gate at every stage.

Assets, lighting and polish can hide a dead core loop. This routing skill requires a playable greybox first and proof at every handoff.

View skill on GitHub
3d-asset-qualityProduction gates

Require functional construction, thickness, bevels and geometry-only acceptance renders.

Textures can disguise primitive geometry. This gate makes every asset explain what it is, how it is built and how it is used with materials disabled.

View skill on GitHub
asset-judge-loopProduction gates

Let independent judges score multi-asset work; builders never certify their own output.

Large batches dilute effort and self-review hides weak assets. Fresh judges score the minimum, reconcile inventory and send failures back to rebuild.

View skill on GitHub
verify-by-playingProduction gates

Accept gameplay only after real input, route traversal, camera ownership and exercised verbs.

Green tests can miss dead controls, sealed routes and hijacked cameras. This gate proves the shipped build by actually playing it.

View skill on GitHub
02

Cross-model subagents

2 skills
codex-subagentCross-model subagents

Run Codex as a hardened second-model reviewer or structured parallel fanout from Claude Code.

Ad hoc second-model calls lose context, structure and durable evidence. The wrappers enforce safe defaults, schemas, timeouts and recorded run artifacts.

View skill on GitHub
grok-subagentCross-model subagents

Run Grok as an independent reviewer, structured judge or parallel fanout from Claude Code.

A third opinion is useless when the call is unbounded or unverifiable. The wrappers preserve structured output, cost, effort and resumable run evidence.

View skill on GitHub
What is a skill?

Rules earned the expensive way.

01 · The failure

AI can produce plausible work and call it finished.

02 · The gate

Each skill encodes one hard-won production rule.

03 · The receipt

Acceptance comes from manifests, renders, independent judges and real play—not confidence.

Open production method

Install RemakeBench Skills v0.1.

Use them, adapt them, or submit a focused improvement. The catalogue and its source remain public under the MIT licence.

View repository Read CONTRIBUTING.md
Claude Code · terminal
claude plugin marketplace add RemakeBench/skills
claude plugin install remakebench-skills@remakebench

Builder is the project-specific layer: exact benchmark prompts, released source projects, RemakeBench Harness, implementation material, troubleshooting, progress and support. The skills stay public. Explore Builder →