Stake Engine Skills for Claude Code

Take a developer from a blank folder to a working slot game on Stake Engine — Claude does the cloning, the version pinning, the smoke test, and answers every question along the way.

Two Claude Code skills bundling the official Stake Engine Math SDK and Web SDK workflows. Drop them in, ask Claude "I want to start a Stake Engine slot game", watch a sample run.

What's inside

Skill	Bootstraps	Then helps with
stake-math-sdk	Clones StakeEngine/math-sdk, Python 3.12+ venv, runs a sample	GameState, bet modes, calculations (lines, ways, cluster, scatter, tumble, board), the optimizer, books + lookup-tables, math approval (RTP 90-98%, ±0.5% between modes)
stake-web-sdk	Clones StakeEngine/web-sdk, Node 22.16 via nvm, pnpm 10.5, smokes Storybook	Monorepo navigation, book events, Pixi components, XState context, RGS HTTP API, frontend approval (static-only, CDN-only, mini-player, mobile, autoplay rules)

Both skills point at upstream Stake Engine repos as the source of truth — the official docs at https://stakeengine.github.io/math-sdk/ and the repos in https://github.com/stakeengine. They stay in sync via scripts/update-sdk.sh.

Install

# Download the latest release tarball, then:
tar -xzf stake-engine-skills-v0.1.0.tar.gz
cd stake-engine-skills
./install.sh

That copies both skills into ~/.claude/skills/. Restart Claude Code (or open a new session) to load them.

Or clone the repo directly:

git clone https://github.com/ReSkin-Games/stake-engine-skills.git
cd stake-engine-skills
./install.sh

install.sh will offer to also install Stake's official MCP server for live doc access (stake-engine-docs). Recommended — accept with [Y]. Skip with SKIP_MCP=1 ./install.sh if you'd rather install it manually later via ./install-mcp.sh.

Use

Open Claude Code in any folder and say:

"I want to start a Stake Engine slot game."

The math skill activates, runs detect-state.sh, finds no SDK present, and walks through the bootstrap. After it finishes there's a working math-sdk clone next to you, the venv is set up, and a sample game has produced its books_*.jsonl and lookUpTable_*.csv.

When switching to frontend:

"Now let's set up the Web SDK and start playing the sample in Storybook."

The web skill activates, clones web-sdk, installs Node 22.16 via nvm, runs pnpm install, and tells you to run pnpm run storybook --filter=lines.

After that — ask the skill anything about the docs, the architecture, the approval checklist, how to add an event, how the optimizer works. The skills carry distilled references for every page of the official docs.

How it works

Each skill follows the Agent Skills open standard:

stake-math-sdk/
├── SKILL.md              # frontmatter + glossary + navigation, ≤500 lines
├── references/           # 27 on-demand reference files (the docs, distilled)
│   ├── bootstrap.md      # the walkthrough
│   ├── gamestate.md      # GameState, MRO, run_spin
│   ├── optimizer.md      # RTP targeting
│   ├── approval-math.md  # what gets enforced on approval
│   ├── rgs-api.md        # full HTTP contract
│   └── ...
└── scripts/              # atomic shell scripts, each does one thing
    ├── detect-state.sh   # routes to the next needed step
    ├── clone-sdk.sh      # idempotent clone-or-update
    ├── install-python-env.sh
    ├── smoke-test.sh     # runs sample, verifies books output
    └── update-sdk.sh

stake-web-sdk/ follows the same shape with 21 references and 5 scripts pinned to Node 22.16 + pnpm 10.5.

The SKILL.md is loaded into Claude's context when the skill activates; references are read on-demand by topic; scripts are executed (their output enters context, not their source). This is progressive disclosure — the user gets the right material at the right depth.

Requirements

	Math SDK skill	Web SDK skill
OS	macOS or Linux (Windows: use WSL)	macOS or Linux (Windows: use WSL)
Git	required	required
Other	Python 3.12+, make, optionally Rust/Cargo for the bundled optimizer	Node 22.x, pnpm 10.x (install steps included in bootstrap)

The bootstrap scripts check prerequisites and tell you what to install if anything is missing.

Supported tools

The skills are primarily authored for Claude Code (full progressive disclosure, bootstrap automation, on-demand references). Other AI dev tools get a lighter integration:

Tool	Skill activation	Bootstrap automation	Live doc search
Claude Code	full skill (SKILL.md + 21–27 references) auto-loads by trigger	atomic scripts run by the agent	MCP server (auto-registered by install.sh)
Cursor	AGENTS.md at repo root (passive context)	scripts available, agent must invoke explicitly	MCP server (manual add to ~/.cursor/mcp.json)
Windsurf	AGENTS.md (passive)	scripts available, agent must invoke explicitly	MCP server (manual add to ~/.codeium/windsurf/mcp_config.json)
Codex CLI	AGENTS.md (its native convention)	scripts via tool calls	MCP server (manual add)
Aider / Copilot / others	manual: paste AGENTS.md into instructions	none — run scripts by hand	MCP only if the tool speaks MCP

For non-Claude-Code tools:

1. Copy AGENTS.md to the root of your project (or set it as a global rules file).
2. Run ./install-mcp.sh to set up the live-doc MCP server. The script prints the JSON snippet to paste into your tool's MCP config.
3. Bootstrap scripts in stake-math-sdk/scripts/ and stake-web-sdk/scripts/ can be run from any shell — you just don't get the auto-activation that Claude Code provides.

If you'd like a first-class adapter for a specific tool (Cursor .mdc rules, Windsurf workflows, etc.), open an issue or start a discussion.

Updating

To pull fresh SDK code:

~/.claude/skills/stake-math-sdk/scripts/update-sdk.sh <path-to-math-sdk>
~/.claude/skills/stake-web-sdk/scripts/update-sdk.sh <path-to-web-sdk>

To get a newer skill bundle: re-extract the new tarball or git pull and run ./install.sh again.

Building a release

For maintainers:

# Bump VERSION + update CHANGELOG.md, then:
./pack.sh
# Produces ../stake-engine-skills-v<VERSION>.tar.gz

Source of truth

All references in these skills were distilled from:

• StakeEngine/docs — the open-source docs site (cloned at build time)
• The upstream README files of StakeEngine/math-sdk and StakeEngine/web-sdk
• Stake's published Approval Guidelines

If you spot a divergence between the skill and current Stake docs, please open an issue. Versions and changes are tracked in CHANGELOG.md.

License

MIT — see LICENSE.

The Stake Engine SDKs themselves are governed by their own upstream licenses.

Acknowledgements

Built by ReSkin Games. The Stake Engine SDKs are by Stake and their authors — full credit and thanks for keeping the platform open.

Skills authored with Claude Code, following Anthropic's skill best-practices.