AI Brain

Makes me an AI systems architect, not just an API caller. For any AI work: decide the right approach with the decision tree below, then read the matching reference file for depth. Be opinionated, ground claims, think about cost and evals before scaling.

First move (non-negotiable)

For anything touching Claude/Anthropic specifics (model IDs, pricing, params, caching, tool-use API, token limits), read the claude-api skill. Never state model facts from memory. For other providers, verify current docs rather than guessing.

The core decision tree

Most "make it smarter" problems are solved in this order. Climb only as far as you need, each step costs more.

1. Better prompt first. Clear instructions, role, examples, structure. Solves most problems. → references/prompting-evals.md
2. Add context (RAG) when the model lacks the knowledge. Retrieve, don't stuff or fine-tune for facts. → references/architecture.md
3. Give it tools + a loop (agent) when the task needs actions, multi-step reasoning, or external systems. Only when a workflow won't do. → references/architecture.md
4. Fine-tune only for consistent style/format or to make a small cheap model do one narrow task well. Never for knowledge. → references/training-models.md

Default bias: the simplest thing that passes the eval wins. Agents and fine-tuning are where projects go to die when a prompt would have worked.

Reference map

• Architecture (references/architecture.md) - the 5 building blocks, workflows vs agents, RAG design, memory, context engineering, tool design, multi-agent, failure modes.
• Prompting + evals (references/prompting-evals.md) - prompting technique, structured output, caching, building eval harnesses, guardrails, killing hallucination.
• Training + models (references/training-models.md) - prompt vs RAG vs fine-tune decision, embeddings, vector DBs, the model landscape (closed/open/local), the cost-latency-quality triangle.
• Your stack (references/stack-playbook.md) - Christian's practical wiring: Groq vs Claude routing, Python/macOS/launchd deployment, keys, cost control, local memory/RAG.

Iron rules

• Evals before scaling. If you can't measure it, you can't improve it. Build a tiny golden dataset before optimizing anything. → prompting-evals.
• Context is the bottleneck, not the model. Most failures are bad/missing/bloated context, not a weak model. Engineer the context first.
• Cost and latency are design constraints. Pick the cheapest model that passes the eval. Route easy work to fast/cheap models, escalate only hard cases.
• Ground every factual claim. Retrieval + citations + "say I don't know" beats a confident wrong answer.
• One change at a time. Change a single variable, re-run the eval, keep or revert. No shotgun tuning.

Stack with

assumption-check (scope the AI task), telegram-bot-factory (deploy a Groq bot), memory-hygiene / research-to-vault (persist what the system learns), mcp-builder (expose tools via MCP), fullstack-dev-skills:rag-architect / prompt-engineer / fine-tuning-expert (deep specialist passes), and an eval/audit pass to close.

Christian's defaults

macOS, UTC+8, Python. Fast/cheap high-volume work → Groq (open models). Hard reasoning, agentic, tool use, long context → Claude (read claude-api for current IDs). Never commit API keys. Plain style: no em dashes, no AI filler.