ysz7

build-agent-loop

Builds a production-shaped agent tool-use loop with step budgets, tool dispatch, error handling, truncation, loop detection, and tracing. Use when the user asks to build an AI agent, create an agent from scratch, set up a tool-use loop, add tool calling to an LLM, make a model that can use tools, or scaffold an agentic backend. Also use when an existing agent needs governance added - step caps, cost limits, or structured traces.

build-eval-set

Creates a labelled evaluation dataset and scoring harness for an LLM system - sourcing cases from real usage, categorising them, adding adversarial and should-fail cases, and wiring deterministic scorers plus a baseline. Use when the user wants to evaluate an LLM system, measure prompt or model quality, build a test set for AI, set up eval CI, or prove that a change actually helped.

build-rag-pipeline

Builds a production-shaped retrieval-augmented generation pipeline - parsing, chunking with metadata, embedding, hybrid retrieval, reranking, grounded answering with citations and abstention. Use when the user asks to build RAG, add document search, make a chatbot over their docs, index a knowledge base, do semantic search, or let a model answer from a corpus it was not trained on. Also use when an existing RAG system needs hybrid search, reranking, or evaluation added.

design-agent-tools

Designs or repairs the tool interface an agent programs against - names, descriptions, JSON schemas, response formatting, and error messages. Use when an agent picks the wrong tool, ignores a tool, passes bad arguments, blows up its context on tool output, or repeats calls. Also use when adding new tools, converting REST endpoints into agent tools, or reviewing an MCP server's tool surface.

diagnose-rag-failure

Localises a wrong RAG answer to a specific stage - retrieval miss, ranking failure, or generation failure - by checking each stage against evidence rather than guessing. Use when a RAG system returns a wrong or incomplete answer, hallucinates despite having documents, says it cannot find something that exists, cites the wrong source, or when retrieval quality regressed after a change.

write-agent-tests

Writes the test suite for an AI agent - deterministic unit tests with mocked tools, contract tests for tool schemas, trajectory tests for multi-step behaviour, adversarial tests for injection and abuse, and a small graded eval set. Use when the user asks to test an agent, add tests for LLM code, set up CI for an agent, check agent reliability, or when an agent regression shipped unnoticed.

write-system-prompt

Writes or restructures a production system prompt with a stable cacheable prefix, positive constraints, a numbered procedure, an explicit output contract, boundary examples, and a stop condition. Use when the user asks to write a prompt, improve a prompt, fix a model that ignores instructions, reduce prompt cost, make outputs more consistent, or turn ad-hoc instructions into something production-grade.

blueprints

Architectural blueprints this project has already settled on, covering agent core, agent protocols, agent workflows, data pipelines, evaluation, LLM infrastructure, MCP servers, prompting and retrieval. Use when building, extending or fixing a feature in one of those areas, so the work follows the recorded approach and its reasoning instead of being designed again from scratch. Not for topics with no blueprint here, and not for ordinary coding questions.