relevance-builder-kit

Auto-generate test cases and run platform evals on a Relevance AI agent. Covers golden sets, quick smoke tests, publish gates, and ongoing performance monitoring of production conversations. Use when the user says "test this agent", "evaluate this agent", "set up evals", "run an eval", "create test cases", "set a publish gate", "add a golden set", "monitor production quality", or asks how to validate before going live. Quick mode (3 cases, auto-fix loop) for after-change smoke tests; Full mode (5-8 cases, user-reviewed) before launch or major redesign.

Design and build a clean starter agent for any new use case. Use this skill when starting a new build, when someone asks "how should I build my first agent for X?", "what goes in a v1 agent?", "how do I avoid ending up with a generic stock-template-looking agent?", or when sanity-checking before shipping. Covers the 12-point design rubric, architecture layers (variables vs knowledge tables vs tools vs placeholders), the v0 to vN roadmap pattern, build-fresh-not-fork principle, system prompt anti-patterns, tool selection rules, smoke test via async pattern. Pairs with `/agent-build-patterns` for general design philosophy.

Diagnose a Relevance AI agent or workforce for config, prompt, tool, and credit-spend issues, then recommend ranked optimizations with concrete prompt rewrites and credit-saving estimates. Use when the user says "optimise this agent", "audit this workforce", "review this agent's prompt", "find waste in this build", "why is my agent slow / expensive / unreliable?", "the agent is using too many credits", or asks for a health check on an existing build. Reads platform state via native MCP tools (no writes without permission); works on any connected project.

Design philosophy, architecture patterns, and decision guides for production-grade Relevance AI builds. Use this skill when designing a new agent or tool, reviewing existing architecture, debugging production failures, deciding where logic should live (code vs LLM vs knowledge table), structuring multi-agent workforces, eliminating duplicated tools, adding audit trails, or when someone asks "is this the right pattern?" or "why is my agent doing X wrong?". Covers Unit of Action, Note Step, Centralize Rules, Code Over LLM, Parameterize, Compute at Right Level, Data Over Routing, and Audit Enrichment patterns.

Review a build or debug session, extract reusable learnings, and update the knowledge base or share with the wider community via PR. Use this skill at end of session, when the user says "what did I learn", "save this pattern", "document this discovery", "let's do a retro", "capture this before I forget", "update the knowledge base", or "share this back". Also use when someone wants to graduate a discovery into BUILD_PRACTICES.md, PLATFORM_MECHANICS.md, or agent-build-patterns.

Capture a single mid-flow insight as a well-scoped repo PR. Highly opinionated - refuses ~half of invocations on a substance bar (actually happened, non-obvious, clear canonical home, no contradictions, in scope, reusable). Use when the user says `/improve`, says `go` immediately after Claude surfaced an "Improvement spotted" line, or says "capture this and PR", "raise this improvement", "PR this learning", "improve the kit with this". Do NOT use for end-of-session retros (use /capture-learning) or build-specific docs (edit builds/{build-name}/...).

First-time setup for this kit. Setup is an interactive bash script (setup.sh) that runs in the user's terminal -- not a conversational skill. Use this skill when the user asks "/setup", says they just cloned the kit, asks how to get connected, or wants to set up MCP / statusline. Tell them to exit Claude Code and run `bash setup.sh` from the kit root.

Document a Relevance AI workforce and all its agents -- fetching config from the platform, creating local markdown docs. Use when the user says "document this workforce", "create docs for the workforce", "write up the workforce", or wants a full documentation package for a workforce and its agents. Also use when someone asks to document agents within a workforce context.