AI & LLM Integration for Claude Code

Researches the business domain and real-world application context of the AI system being built. Surfaces domain expert evaluation criteria, industry-specific failure modes, regulatory context, and what "good" looks like for practitioners in this field — before the eval-planner turns it into measurable rubrics. Spawned by /gsd:ai-integration-phase orchestrator.

Retroactive audit of an implemented AI phase's evaluation coverage. Checks implementation against the AI-SPEC.md evaluation plan. Scores each eval dimension as COVERED/PARTIAL/MISSING. Produces a scored EVAL-REVIEW.md with findings, gaps, and remediation guidance. Spawned by /gsd:eval-review orchestrator.

Designs a structured evaluation strategy for an AI phase. Identifies critical failure modes, selects eval dimensions with rubrics, recommends tooling, and specifies the reference dataset. Writes the Evaluation Strategy, Guardrails, and Production Monitoring sections of AI-SPEC.md. Spawned by /gsd:ai-integration-phase orchestrator.

Presents an interactive decision matrix to surface the right AI/LLM framework for the user's specific use case. Produces a scored recommendation with rationale. Spawned by /gsd:ai-integration-phase and /gsd-select-framework orchestrators.

Execute image generation via the MeiGen MCP server. Delegates all generate_image calls to keep the main conversation context clean. Use for single images, batch parallel generation, or serial workflows.

Visual and design specialist for UI layouts, color systems, design briefs, sprite specs, animation choreography, and AI image-gen prompts. Use PROACTIVELY for any task where the output is a visual specification or design brief. Part of the operator-kit five-agent starter kit.

Background agent that extracts user profiles, contact cards, and behavioral patterns from chat history. Runs as a daemon service every 30 min.

Implements A2A JSON-RPC protocol for communicating between Claude Code and Vertex AI ADK agents, including AgentCard discovery, session management, async polling, and multi-agent orchestration patterns. Use when integrating with deployed ADK agents on Agent Engine or setting up inter-agent workflows. Trigger with \"set up A2A client\", \"connect to an ADK agent\".

Audits and implements AI safety layers including content filtering, PII detection, bias mitigation, and LLM guardrails. Use when securing an LLM application, reviewing safety architecture, or adding compliance controls. Trigger with \"audit ai safety\", \"add safety guardrails\".

Audit LLM spend, identify top token consumers, design model-tiering strategies, and set up cost alerts before budgets breach. Use when AI costs are rising or untraceable. Trigger with "audit AI spend", "optimize LLM costs".

Designs and ships production AI features — LLM integration, prompt engineering, RAG pipelines, evals, and MLOps. Use when you need an AI architecture decision, a prompt-first vs RAG vs fine-tune call, or an eval harness for an existing feature. Trigger with \"build this AI feature\", \"design the RAG pipeline\".

Designs model serving infrastructure, blue/green rollouts, and canary release plans with explicit rollback triggers. Use when you need inference API configuration, a traffic-splitting strategy, or a deployment topology audit. Trigger with \"design the model serving setup\", \"plan a canary release\".

Detects and diagnoses ML model decay in production — data drift, concept drift, and silent degradation. Use when models may have stopped working, distributions have shifted, or monitoring needs a design. Trigger with \"audit my ML monitoring\", \"detect model drift\".

Designs embedding pipelines and vector search systems — model selection, ANN index tuning, hybrid search, and index freshness monitoring. Use when building semantic search, RAG infrastructure, or diagnosing retrieval quality issues. Trigger with \"design embedding pipeline\", \"optimize vector search\".

Builds LLM eval harnesses — benchmark suites, automated regression pipelines, golden set management, and human eval orchestration. Use when measuring model quality, wiring evals into CI, or auditing benchmark validity. Trigger with \"design eval harness\", \"build LLM regression suite\".

Transforms raw data into model-ready features — leakage audits, encoding strategies, feature stores, and reproducible pipeline design. Use when building ML features, auditing for data leakage, or designing a shared feature store. Trigger with \"build feature pipeline\", \"audit features for leakage\".

Selects algorithms, tunes hyperparameters, and builds reproducible training pipelines from baseline to production. Use when choosing a model architecture, designing a tuning strategy, or auditing training code for leakage and reproducibility. Trigger with \"design training pipeline\", \"tune model hyperparameters\".

Provides production-ready code examples from official Google Cloud repos — ADK samples, Agent Starter Pack, Genkit flows, Vertex AI fine-tuning, and Gemini function calling patterns. Use when building AI agents or workflows on GCP and you need battle-tested starter code. Trigger with \"show me an ADK agent example\", \"give me a Genkit starter template\".

Designs and implements production-grade Firebase Genkit flows (Node.js, Python, Go) including RAG, tool calling, model integration, and Cloud Run / Firebase deployment. Use when building or debugging Genkit AI workflows, setting up Gemini model plugins, or deploying agentic pipelines. Trigger with \"create a Genkit flow\", \"design AI workflow with Genkit\".

Designs and audits AI guardrail layers — input/output filters, PII detection, content moderation, and runtime policy enforcement. Use when adding safety controls to an LLM feature or auditing existing ones. Trigger with \"design guardrails\", \"audit our AI safety controls\".

Delivers production-ready LLM API integration patterns including retry/backoff, rate limiting, multi-provider fallback, streaming, and cost tracking across OpenAI, Anthropic, and Google. Use when wiring an LLM into a production service for the first time or hardening an existing integration. Trigger with "integrate an LLM API", "production LLM setup".

Recommends the right LLM model for any task by weighing quality, latency, cost, and context-window requirements across OpenAI, Anthropic, Google, and open-source options. Use when choosing a model for a new feature or reducing spend without sacrificing quality. Trigger with "which model should I use", "pick the best LLM for this".

Designs versioned system prompts, few-shot libraries, and chain-of-thought patterns with A/B testing and regression coverage — treats prompts as production code. Use when engineering a production LLM feature, auditing a prompt library for drift, or building prompt versioning infrastructure. Trigger with \"design this system prompt\", \"audit our prompt library\".

Designs, patterns, and debugs LLM prompts using CoT, few-shot, zero-shot, and meta-prompting techniques, with structured templates and token-efficiency guidance. Use when authoring a new prompt or diagnosing why an existing one produces poor output. Trigger with "design a prompt", "help me write a prompt".

Detects and mitigates prompt injection, jailbreaks, and adversarial input attacks against LLM applications. Use when hardening a system prompt, reviewing LLM input handling, or implementing injection defenses. Trigger with \"defend against prompt injection\", \"harden llm inputs\".

Reduces LLM API costs 50-90% by compressing prompts, selecting the right model tier, and applying caching and batching strategies with measurable ROI calculations. Use when an LLM workflow is too expensive or slow and you need data-driven optimization. Trigger with "optimize this prompt", "reduce my LLM costs".

Implements production-grade Python/FastAPI backends with async patterns, PostgreSQL/Alembic migrations, auth, and LLM integrations strictly from sprint API contract and backend specs, returning a BACKEND IMPLEMENTATION REPORT. Use when building or updating Python API services in a sprint. Trigger with "implement backend sprint", "build FastAPI endpoint".

Designs production RAG pipelines covering chunking strategy, embedding selection, retrieval patterns (basic, reranked, hybrid, multi-query), and evaluation metrics. Use when building a knowledge-grounded Q&A system or improving retrieval accuracy. Trigger with "design a RAG system", "help me build a knowledge base".

Designs retrieval reranking pipelines, relevance scoring systems, and learning-to-rank models with rigorous NDCG/MRR evaluation. Use when ranking quality is poor or a reranker is needed. Trigger with \"improve my search ranking\", \"design a reranking pipeline\".

Optimizes LLM context windows through token budgeting, chunking strategy, and truncation design. Use when you need to control token spend, design a chunking pipeline, or audit token usage in a production AI system. Trigger with \"design my token budget\", \"fix my context overflow\".

Instruments LLM systems with tracing, span capture, cost attribution, and debugging tooling. Use when you need to add observability to AI pipelines, debug model behavior, or audit LLM cost attribution. Trigger with \"instrument my LLM calls\", \"debug my AI system with traces\".

Designs LLM fine-tuning pipelines using PEFT/LoRA, RLHF, and instruction datasets, and systematically optimizes prompts before recommending fine-tuning. Use when prompt engineering alone isn't achieving target quality or you need a smaller, cheaper model for a specific task. Trigger with \"design a fine-tuning pipeline\", \"optimize my prompts\".

Designs embedding pipelines and vector search systems for semantic search, RAG, and similarity applications. Use when you need to build a RAG pipeline, choose a vector database, or audit retrieval quality. Trigger with \"design my RAG pipeline\", \"help me choose a vector database\".

Hippocampus Agent Instructions

Stage 2 clustering worker for the conference-scheduling pipeline. Turns event records into a navigable theme map — 6-8 coarse themes with sub-clusters, an explicit outlier bucket, and soft top-k affinities for every event — using the embed-then-label pipeline (resources/cluster.py) when the libraries are present and an LLM-reasoned hierarchical fallback when they are not. Embeddings group; the cartographer writes the human-readable labels. Reusable across conferences; hardcodes no conference spe

An orchestrating agent that collaboratively helps engineers build graph-based retrieval-augmented generation systems. Guides users through knowledge graph construction, embedding strategy design, retrieval orchestration, system integration, and evaluation. Use when user mentions knowledge graph, GraphRAG, graph retrieval, entity extraction for RAG, Neo4j with LLM, or building graph-augmented AI systems.

An ML/data math intuition coach that explains anything involving vectors, matrices, transformations, or high-dimensional spaces in the style of 3Blue1Brown — geometric first, algebraic second, with an explicit bridge between them. Covers linear algebra, calculus, probability, and the ML primitives built from them (attention, embeddings, PCA, layer norm, gradient descent, diffusion, contrastive loss). Use when user wants intuitive understanding of an ML or math concept, asks "why does X work geom

Weekly ML/systems signal radar for substacker. Scans ~30-50 curated intuition-first sources (Olah, Weng, Karpathy, Jay Alammar, Raschka, Willison, Transformer Circuits, Interconnects, arXiv cs.LG, Hugging Face papers, etc.) for signal-not-noise items, cross-references against topic-ledger, produces a lean digest of ≤10 items. Weekly Friday evening. Not daily. Trigger keywords: trends, ML news, weekly digest, signal, watchlist, external signal, what's new in ML.