Appendix E · One-Page Cheat Sheet

Everything worth remembering, fit on one page if you print it at reasonable margins. Stick it on the wall above your desk.

---

Mental model

An LLM is a stateless function f(text, knobs) -> text. You provide the memory, the retrieval, the tools, the evals, and the humans in the loop.

flowchart LR
    A[LLM] -->|+ memory| B[Chatbot]
    B -->|+ retrieval| C[RAG]
    C -->|+ tools| D[Agent]
    D -->|+ other agents| E[Multi-agent]
    E -->|+ standard interface| F[MCP platform]
    F -->|+ autonomy + evals| G[Agentic software]

---

Default stack (April 2026)

Layer	Default	Alternate
Heavy model	Claude Opus 4.7	GPT-5
Fast model	Claude Haiku 4.5	Gemini Flash 3
Embeddings	voyage-3 or text-embedding-3-large	bge-large-v2 (self-hosted)
Vector DB	pgvector on Postgres	Qdrant / Weaviate / Pinecone
Orchestration	LangGraph (Python) / Spring AI (Java)	Temporal + LLM calls
Observability	Langfuse	LangSmith / Helicone / OTel
Evals	promptfoo + pytest	LangSmith / Braintrust
Coding	Claude Code + Cursor + Copilot	Zed AI / GitHub Copilot Workspace
Local model	ollama with qwen2.5-coder	LM Studio
Protocol	MCP for tools and resources	OpenAPI for classical APIs

---

Decision tree

flowchart TB
    A[Problem] --> B{Needs fresh facts?}
    B -- yes --> C[RAG]
    B -- no --> D{Needs tone / format?}
    D -- yes --> E{Few-shot fixes it?}
    E -- yes --> F[Few-shot prompting]
    E -- no --> G[Fine-tune LoRA]
    D -- no --> H{Needs deep reasoning?}
    H -- yes --> I[Reasoning model + CoT]
    H -- no --> J{Needs to act?}
    J -- yes --> K[Tool use / agent]
    J -- no --> L[Just call the model]
    K --> M{Irreversible?}
    M -- yes --> N[Human checkpoint]
    M -- no --> O[Full auto]

---

The five things you always do

1. Set max_tokens. Runaway generations cost money.
2. Read every diff. Never merge code you don't understand.
3. Measure with evals. Don't change a prompt without a test set.
4. Log tokens, latency, cost. On every call, from day one.
5. Gate destructive actions. Human in the loop on writes, deploys, money.

---

Prompting heuristics

• Put instructions and examples in the system prompt. Put data in the user prompt.
• Use XML tags to delimit sections: <context>...</context>, <question>...</question>.
• Ask for structured output. Validate with a schema. Retry on parse error with the error message.
• Use chain-of-thought ("think step by step") for reasoning tasks on non-reasoning models.
• Use self-consistency (sample N, vote) when the task is high stakes.
• Keep the system prompt under 1k tokens unless you have a reason; longer usually hurts.

---

RAG checklist

• Chunk at 500–1,000 tokens, with 50–150 overlap.
• Use hybrid search (vector + BM25) unless you have a reason not to.
• Rerank the top 20 → top 5 with a cross-encoder.
• Store metadata for filtering (tenant, date, doc_id, permissions).
• Return citations with every answer.
• Evaluate retrieval separately from generation — they fail differently.
• Consider contextual retrieval (prepend chunk with a summary before embedding).

---

Agent checklist

• One clear goal per agent.
• A short list of well-described tools. Descriptions are prompts.
• Budget on tool calls and tokens. Hard stop.
• Human confirmation for destructive actions.
• Treat tool outputs as untrusted input (prompt injection).
• Observability on every step. Traces, not logs.
• Evals on multi-turn scenarios, not just single turns.

---

Cost levers (largest first)

1. Use a smaller model where it works (Haiku / Flash / Mini).
2. Cache long system prompts (Anthropic, Gemini prompt caching).
3. Shorten inputs: better RAG beats bigger context.
4. Parallel tools reduce wall-clock latency, not cost — but UX matters.
5. Batch non-interactive jobs overnight at discounted rates.
6. Distill to a smaller model for very high-volume workflows.

---

Daily practice

flowchart LR
    M[09:00 Triage
agent summarizes] --> P[09:30 Plan
agent + you]
    P --> D[10-12 Deep work
pair coding]
    D --> R[14-16 Review diffs
draft PR text]
    R --> U[16-17 Unblock + learn
20 min paper]
    U --> S[17:30 Shutdown
tomorrow's top-3]

---

Skills that compound

• Taste. Judgment. Systems thinking.
• Writing specs, docs, prompts, decisions.
• Reading diffs with genuine skepticism.
• Teaching juniors and peers.
• Staying curious.

Skills that decay

• Rote syntax recall.
• Writing boilerplate.
• Framework-of-the-month chasing.
• Memorizing APIs you call twice a year.

---

Emergency commands

# Install the CLI
npm install -g @anthropic-ai/claude-code
claude

# Run a local model offline
ollama run qwen2.5-coder:14b

# Test a prompt with evals
promptfoo eval

# Tokenize quickly
python -c "import tiktoken; enc=tiktoken.encoding_for_model('gpt-4'); print(len(enc.encode(open('file.txt').read())))"

# Cheapest smart model for a one-off question
claude -p "your question here" --model claude-haiku-4-5

---

The one rule

Ship something this week.

The saga is still being written. Go write a chapter.

---

Back to index · Preface · Chapter 1

← Table of Contents