Using slotplate with AI agents
Point any coding agent at your slotplate project and keep it on-spec.
slotplate is designed to be edited by AI agents as a first-class case. Every principle has a lint rule. Every common task has a slash command. Every concept has a docs page the agent can fetch.
The feeds
-
/llms.txt— terse machine-skimmable index, one line per doc page. Good for routing queries. -
/llms-full.txt— the entire manual concatenated. One fetch, everything. Paste into your agent at session start.
Both regenerate on docs build. Point your agent at
https://slotplate.dev/llms-full.txt once per session and it has the full manual.
Claude Code
CLAUDE.md is committed to the template root. It codifies the 10 principles
as project rules. Slash commands live in .claude/commands/:
/add-symbol Cherry # scaffolds a new ReelSymbol + tests + registration
/add-phase BonusSpin # scaffolds a new FSM phase
Open the project in Claude Code and the rules apply automatically.
.claude/settings.json pre-allows the common commands (pnpm,
git status, etc.) so you're not clicking "approve" for every step.
Cursor, Codex, Aider, Continue
AGENTS.md mirrors the principles for non-Claude agents. Point your agent at:
AGENTS.mdin the project root, orhttps://slotplate.dev/llms-full.txtfor the full manual.
Cursor users: symlink AGENTS.md → .cursor/rules/main.md and
it loads automatically as a rule.
Recommended agent prompt
You are editing a slotplate project — an opinionated slot-game CLIENT
boilerplate on pixi-reels. This is a client, not a server: no win
evaluation, no paytable, no RTP here. The server returns resolved
results; the client renders them.
Read CLAUDE.md before making any changes. The 10 principles listed
there are enforced by Biome; diffs that violate them will not pass.
Key rules:
- src/domain/ holds WIRE TYPES only — no evaluators, no pure math.
- Use Ticker.schedule, never setTimeout.
- Only src/view/ and src/presenters/ import pixi-reels.
- Every FSM phase is a file in src/flow/phases/.
- State mutates through MobX @action methods only.
For anything not covered here, fetch
https://slotplate.dev/llms-full.txt. What the agent rules prevent
- Creating a
src/domain/winEval.tsbecause the agent "saw it in another slot repo." - Using
setTimeoutfor win-display hold timing. - Importing
pixi.jsfrom a phase or store. - Silent
catcharound network errors. as anyto silence the type system without a comment.- Writing a paytable JSON to "make the mock more realistic."
These are the failure modes of AI-edited slot code. slotplate encodes each as a lint rule, a docs page, and a principle the agent reads at session start.
Building on top
Add your own slash commands in .claude/commands/. Mirror them in
AGENTS.md so non-Claude agents know about them. The pattern:
- Markdown file with a frontmatter
descriptionandargument-hint. - Body is a checklist the agent works through.
- End with "run
pnpm typecheck && pnpm test && pnpm lint" to loop on green.