Heidelberg AICurriculum

Canvases

27 diagrams — one per chapter. The architecture of a topic in a single picture: what talks to what, where a model sits, and where you come in. Each one links to the chapter that teaches it.

Chat with AI1

Which chatbot for the job? A decision fork from "what do you need?" to ChatGPT, Claude, Gemini and Perplexity, each labelled with the task it fits. Which chatbot for the job? four frontier assistants — pick by the task, not the brand What do you need? ChatGPT all-rounder — Q&A, drafts, code Claude careful writing & analysis Gemini Google-grounded · multimodal Perplexity research with cited sources anything write & analyse images / search research Same question, four strengths — no setup, just open and prompt.
Frontier chatbots

Better prompting4

The context window Four moves — Write, Select/RAG, Compress, Isolate — all feed into the finite central context window that the model sees each turn. The context window what the model sees — managed with four moves context window what the model sees this turn finite — tokens cost, and rot Write put what matters in Select · RAG pull in only the relevant Compress more signal, fewer tokens Isolate separate concerns & agents Most agent failures come from poor context — not a weak model.
Context Engineering
Anatomy of a loop you design the system that prompts the agent — not each turn 1 · Trigger schedule · event · you 2 · Act agent does one step 3 · Check tests · reviewer pass ✓ goal verified fail → fix & retry stop after N tries · budget three rungs — hand off more of the driving 1 · Closed loop verify every change 2 · Unattended Ralph + anchor files 3 · Orchestrate batch · subagents the whole technique, in one line: while :; do cat PROMPT.md | claude-code ; done Trigger · Act · Check · repeat — until a verified goal, with a way to stop.
Loop Engineering
Anatomy of a strong prompt A vague prompt vs. a structured prompt with Role, Context, Task, Examples, Format and Constraints — the six parts a model can act on. Anatomy of a strong prompt say what you want — then shape the response ✕ vague "write about dogs" → generic, guessy ✓ a prompt it can act on Role You are a vet writing for first-time owners. Context Audience: UK, no prior experience. Task Explain crate training as 5 concrete steps. Examples Match this tone: "short, upbeat sessions…" Format Numbered list, ≤150 words, plain English. Constraints No jargon. Say so if you're unsure. + advanced — let it think step by step · ground in your sources Same ask — a vague line vs. the parts a model can actually act on.
Prompt Engineering
The test-first prompt loop Write input → expected examples first, then loop Red → Green → Refactor, and re-run the whole set on every prompt or model change to catch regressions. The test-first prompt loop decide what "right" looks like before you write the prompt — then let it drive every edit 1 · Write the examples first a handful of input → expected pairs — edge cases & known failures = your checkable spec start red RED run the prompt on your set — a case comes back wrong GREEN make the smallest prompt change that passes it REFACTOR tighten & shorten — every example stays green loop — repeat for the next failing case Keep the file. Re-run the whole set on every change. a prompt edit · a model swap · a silent model update — catch the regression before your users do TDPE is the loop. How you grade each output — code · LLM-judge · human — lives in Evals & testing.
Test-Driven Prompt Engineering

Build apps1

From a sentence to a working app A four-step pipeline — describe, AI builds, preview & refine, publish — with a branch showing how Base44, Lovable and v0 differ in what they hand you. From a sentence to a working app vibe coding — describe it, get a real app, no code Describe plain English prompt AI builds it generates the whole app Preview & refine see it, ask for changes Publish share a live URL what you get differs by tool ↓ Base44 full-stack, backend baked in Lovable you own the code v0 UI only Same describe → build loop; the tool decides what you walk away with.
AI app builders

Automation Tools2

Anatomy of an automation workflow Trigger to AI or transform node to an IF branch that forks into an action on match or an else path, with a dashed error-and-retry loop under the AI node. Anatomy of a workflow trigger → nodes that process & route → action — the n8n / Make / Zapier canvas Trigger webhook · schedule · app event AI / transform summarize · classify · extract fields IF branch route on a condition Action ✓ Slack · email · update DB Else log · skip · stop match no match on error → retry, then alert Each box is a node you wire on the canvas — no code between the boxes.
Automation tools
Anatomy of the extensibility layer Agent core at centre, five plug-ins around it: Skill (top), Slash command (mid-left), Subagent (mid-right), Hook (bottom-left), MCP tool (bottom-right). Anatomy of the extensibility layer an agent is a small core — these five plug-ins give it new powers The agent model + small core — bolt powers onto it Skill SKILL.md loaded on task match reusable · portable across tools Slash command saved prompt, invoke by name shortcut, e.g. /review Subagent helper in its own context offload noisy / parallel work Hook script that fires on an event automatic — no prompt needed MCP tool connects an external system GitHub · database · calendar Same five pieces in every agent — learn them once, carry them everywhere.
Skills, tools & extensions

Agent frameworks1

An agent on a canvas — grounding an LLM, and a multi-agent crew Pattern one grounds an LLM with context from any retrieval method then reasons with tools and memory. Pattern two has a manager delegate to researcher, writer and reviewer agents that merge into a result. An agent on a canvas Flowise · Langflow · CrewAI — wire context, an LLM, tools, and other agents ① GROUNDING AN LLM — give it the right context, then let it reason User asks a question Get context pull in what's relevant LLM reasons tools memory Answer grounded context can come from — pick any, or combine: docs pasted in — no search, simplest vector store — semantic top-k (classic RAG) keyword / full-text search live tool — web search · API · database ② MULTI-AGENT CREW — a manager delegates to specialist agents Manager plans & delegates Researcher gathers sources Writer drafts the answer Reviewer checks & fixes Merged result the crew's output
Build agents

Personal AI Assistants2

How Claude Cowork works Four-step loop: Describe → Review plan → Claude works → Finished work. Inside step 3: parallel sub-agents, isolated sandbox, your tools (Drive/Gmail/Chrome/plugins). How Claude Cowork works describe an outcome → get back finished work (no terminal) 1 · Describe outcome you want + guardrails & folder 2 · Review plan Claude proposes steps you approve first 3 · Claude works files & apps watch & redirect 4 · Finished work report, sheet, folder gaps flagged for you Inside "Claude works" Parallel sub-agents splits a big job, runs pieces at once Isolated sandbox only folders you allow — nothing else touched code runs in a separate VM Your tools Drive · Gmail · Chrome + job-function plugins Finance · Sales · Research… You own the what — outcome and guardrails. Claude works out the how. (/schedule repeats weekly.)
Claude Cowork
Anatomy of a personal assistant You to a chat app to an always-on assistant that uses web, memory and tools, then returns results to you. Hermes and OpenClaw are the ready-made options. Anatomy of a personal assistant install once → chat from anywhere → it works 24/7 You ask in words Chat app Telegram, WhatsApp · web Assistant always on, remembers you, runs while you sleep Done work sent to your chat browse the web memory run tools Uses these to do the job — you use it, you don't build it. Ready-made: Hermes (self-hosted) · OpenClaw (chat-driven)
Personal AI Assistants - Intro

Chat with your Data - RAG1

The RAG pipeline A left-to-right flow: your documents are chunked and embedded into a vector store; a question retrieves the top-k chunks, which hybrid search and reranking narrow to the best few; the model answers from those sources with citations; and the retriever and answer are evaluated for precision, recall and faithfulness. The RAG pipeline answer from your own files — with citations you can check Your docs PDFs · policies KB · runbooks 1 · Chunk & embed split with overlap add chunk context Vector store pgvector · Chroma · Qdrant 2 · Retrieve & rank retrieve top-k chunks vector + BM25 hybrid rerank → top 5–10 place best at start/end beats lost-in-middle 3 · Grounded answer answer from sources only cite each chunk say if not found → cited answer Evaluate — prove it works retriever: context precision · context recall answer: faithfulness (supported claims / total) Under ~200k tokens? Put the docs in the prompt. Bigger / changing corpus → retrieve.
Chat with your own documents

Local AI & Private Cloud3

One URL, open models at cloud speed Your code to a base URL plus API key, forking to Groq, OpenRouter and Cerebras, with fast tokens returned. All OpenAI-compatible, so swapping is one line. One URL, open models at cloud speed point your code at a provider — switch in one line Your code OpenAI-compat call One URL + key base_url + api_key swap = 1 line Groq fastest · LPU chips OpenRouter 200+ models · one key Cerebras wafer-scale · 1M free/day fast tokens back All speak the OpenAI API — pick for speed, model choice, or a free tier, and switch any time.
EU-sovereign inference
Where a local model runs — and what never leaves Your machine holds a downloaded GGUF model, a local runtime, and the app that calls it on localhost; the cloud API stays outside the privacy boundary and receives no data. Where it runs — what never leaves a local model: prompts and data stay on your machine YOUR MACHINE — nothing crosses the dashed line Download a model one .gguf file, once Qwen3.6-35B-A3B 22 GB gpt-oss-120b 63 GB DeepSeek-R1 671B 404 GB Q4 GGUF — fits your RAM Local runtime Ollama :11434 LM Studio :1234 Your app calls localhost OpenAI-style API ✕ no data sent Cloud API OpenAI · Anthropic not in the loop No internet needed, no per-token cost — the trade is your own RAM, disk and GPU.
Local & Private Models
How a change reaches users test on a private copy first — promote only what you have checked 1 · Edit on staging branch — not on prod 2 · Staging private · :3012 safe to break 3 · Verify does it look right? check the real page 4 · Production live · public · users keep it boring push merge not right? fix on staging — never touch prod what actually promotes it (this site): git push origin staging → verify http://…:3012 → merge staging → main → live two copies of the same app — one to break, one to trust Staging — the rehearsal private · disposable · no users watching Production — the show public · live · only proven changes Edit · Stage · Verify · Promote — the same change, proven before it goes live.
Self-host on a private cloud

Write code2

How a coding agent works — read, edit, run, fix, repeat A loop: read context, propose edit, run it, read output; fail loops back to propose edit, pass goes to done. A chatbot stops after propose edit. How a coding agent works reads code, runs it, fixes its own mistakes — not just a chatbot Read context repo · open files · grep Propose edit write & patch the code Run it tests · build · run command Read output pass or fail? errors · diffs · logs fail → fix & retry pass ✓ Done change verified A chatbot stops after "Propose edit". An agent runs it and checks itself.
AI coding assistants
From idea to build plan write the blueprint before any code is written One-liner idea, one sentence + "ask me 5 Qs" The spec · PRD.md user stories · data model screens · acceptance criteria constraints & guardrails = source of truth Build plan ordered steps, you approve Build AI writes code, step by step change? update spec first scope v1 · what it will NOT do the "must-never" rules Catch a wrong assumption in a 10-line plan — not in a finished, broken app.
Product requirements & spec engineering

Building complex codebases1

Anatomy of the AI layer Your git repo contains your code plus the AI layer (CLAUDE.md, commands, skills). Every task runs a Plan → Implement → Validate loop; pass = shipped, fail = fix and retry, which feeds improvements back into the AI layer. Anatomy of the AI layer a second codebase, checked in next to your code, that teaches the agent your project your git repo Your code src/ · tests · the app itself — what you ship + The AI layer CLAUDE.md — lean project rules .claude/commands/ — shortcuts skills · subagents · MCP — on demand checked in together — agent context is reviewed and improved like code drive every task with one loop 1 · Plan fresh-context blueprint 2 · Implement follow the plan 3 · Validate checks must pass pass ✓ shipped fail → fix & retry every miss → new rule in the AI layer ↑ It's context, not a smarter model — give the agent the layer and it stops guessing.
Building complex codebases

AI memory systems1

Marvin — what writes, what reads, what deletes, and what compounds Capture, hybrid index, recall into the prompt, and the promotion loop that turns an expiring conversation into a reviewed lesson. Marvin — a memory that outlives the session what writes · what reads · what deletes — and the one edge that makes it compound CAPTURE — every session Prompt hook your message + its project Answer hook the reply, in chunks fire-and-forget — a memory write never delays a prompt Embedding endpoint text → 768 numbers, local stored chunks and each question, tagged apart The index hybrid: keyword + meaning blended half and half reviewed lessons grown by promotion chatter + open tickets hundreds of lines a day RECALL — every prompt Top 5 passages reserved slots per tier each passage capped tier + date shown, always Into your prompt before the model reads it recall@5 gate — the build fails below its floor, so a dead retriever stays visible LIFECYCLE — what expires, and what is kept before it does Expire — past a weekend, not before verified, backed up, deleted Distill candidates a local model proposes lessons worth keeping Human review approve → durable lesson one click, and revertible Capture and recall are the easy half. Promotion is the edge that compounds: without it you have a short-lived cache; with it, a memory. expiring chatter
Heidelberg Marvin memory

AI + Hardware1

AI across the robotics stack Four layers an engineer drives: write the code, get the model onto the device, the optional LLM/VLM reasoning layer, and the FPGA limit (HLS with AI, not raw HDL). AI across the robotics stack you keep the domain expertise — the assistant speeds up the parts it can You mechatronics & embedded engineer 1 Write the code Claude Code · Cursor · Copilot on C/C++ & ROS 2 review + HIL test 2 Get the model on the device quantize → TFLite-Micro / TensorRT → MCU · Jetson measure first 3 Reasoning layer (optional) LLM/VLM planning + ROS 2 glue · sim-to-real real, not the brain 4 Know the limit — FPGA AI drafts C/C++ HLS you verify — not raw Verilog no un-checked HDL The assistant augments your engineering — it never replaces the control theory, the standards, or the bench.
AI for robotics & edge devices

GOOD TO KNOW4

The eval loop A four-step cycle — Look, Measure, Validate, Improve — repeating on every change, above a grading ladder from cheap code checks to costly human grading. The eval loop stop vibe-checking — grade your AI on real failures, with a number you trust 1 · Look error analysis traces · labels 2 · Measure build graders code/LLM/human 3 · Validate trust the judge vs human labels 4 · Improve fix prompt then re-measure repeat — on every prompt or model change The grading ladder — reach for the cheapest check that works Code-based exact match · schema fastest · reliable LLM-as-judge binary pass / fail for judgment calls Human the gold standard slow · costly cheap → costly · binary by default, numeric only where it earns its keep Real failures decide what to measure — not a prefab metric dashboard.
Evals & testing
Layered defense for an AI app An outer ring of guardrails, monitoring and red-teaming wraps your application. Inside, a trust boundary separates your instructions from untrusted input — the user message, retrieved documents and tool output. Below the boundary, model output is encoded before it reaches another system, and tools run with least privilege plus human approval on risky actions. Defend the AI app you build assume injection eventually lands — separate, encode, restrict, and watch guardrails · monitoring · red-team Untrusted input user message retrieved docs tool output label as DATA, not commands trust boundary — screen here System prompt + model instructions stay inside data in labelled blocks no secrets · no auth rules (enforce in code) Output → encode HTML-escape · param SQL strict CSP · schema-validate Least-privilege tools narrow tools · scoped creds human ok on risky actions log every tool call No single layer holds. Bound the blast radius and detect what gets through.
Keep your AI app secure
What leaves your keyboard Your raw stuff passes a strip-first gate that blocks API keys and masks PII before reaching a cloud chatbot; the most sensitive data stays local. What leaves your keyboard before you paste into a cloud chatbot — strip it, control it, or keep it local your stuff 🔑 API keys · passwords ✕ 👤 names · emails · IDs mask 💬 question + safe data ✓ what you type is what you send strip first remove keys mask the PII only safe text passes cloud chatbot trains on your chats by default? Claude no — opt-in ChatGPT yes — opt out Gemini yes — opt out opting out stops training — not retention chats kept a while for safety review too sensitive? keep it local run a local model — nothing leaves your machine → see "Run AI privately" won't risk it — never send Strip secrets · mask PII · prefer opt-in services · keep sensitive data local.
Privacy & safe AI use
From reading the bill to cutting it A three-step journey — Understand the bill (tokens in and out), Trim per call (cap output, less context, right-size the model), Scale the savings (caching ~90% off, batch 50% off, routing) — with a measure-and-attribute loop feeding back to the start. What AI really costs see the token bill — then cut it without losing quality 1 · Read the bill tokens IN vs OUT $/1M input & output output costs 2–8× input estimate before you run 2 · Trim per call cap output length JSON / structured output send less context right-size the model 3 · Scale savings prompt caching · ~90% off Batch API · 50% off model routing / cascade cheap first, escalate rarely then then Measure & attribute log tokens per request tag by feature · model · user attribution shows which lever to pull next — then repeat The levers stack cache the fixed prefix (~90% off) + batch non-urgent work (50% off) + right-sized model (~5× cheaper) — savings compound ~4 characters ≈ 1 token · cost = tokens × $/1M · trim the OUTPUT side first
What AI really costs

Across the course3

Choosing a model — a filters-first method Step 1 filters on data privacy and EU residency, step 2 on cost versus quality, step 3 on the task. Choosing a model a filters-first method — start with your hardest constraint 1 · Must data stay private / in the EU? yes · sensitive data no · ordinary work KEEP IT PRIVATE / IN THE EU GDPR-safe by design Self-host — free & fully offline laptop → Phi-4-mini · gpt-oss-20b 64 GB Mac → Qwen3.6 · Gemma 4 big box → gpt-oss-120b · Llama 4 EU-hosted cloud (managed) Mistral (FR) · Aleph Alpha (DE) 2 · BEST RESULT, COST-AWARE US cloud — use a DPA / EU region Cheap · high-volume Haiku 4.5 · Gemini Flash-Lite GPT nano · DeepSeek V4 Balanced production Claude Sonnet 5 · GPT-5.4 Top quality · hardest tasks Opus 4.8 · GPT-5.5 · Gemini Pro 3 · Then match the model to the task coding · long documents · reasoning · vision · cheap bulk
Choosing a model
Which AI tool? — a decision tree by goal From a root question, goal branches lead to eight tool families, each a chapter, with example tools. Which AI tool? start from your goal — follow it to a family what's the job? start here ask, write, research build software three ways → automate tasks 24/7 personal agent run models myself Consumer chatbots ChatGPT, Claude, Gemini, Perplexity AI app builders Base44 · Lovable · v0 AI coding assistants Claude Code · Cursor · Copilot Agent frameworks Flowise · Langflow · CrewAI Automation tools n8n · Make · Zapier · Dify Personal AI assistants Hermes · OpenClaw Local & private models Ollama · LM Studio · Jan Inference providers Groq · OpenRouter · Cerebras describe my repo custom offline fast API
Tools decision tree
Where your work actually lives Six steps in three zones. The first two are conversation. The middle two — changed files and a saved snapshot — exist only on your own computer. Sending the snapshot away is the step that creates a second copy; after it, the site rebuilds itself. Where your work actually lives one sentence in, a live site out — and the one step that makes it safe ASKING You ask in plain words Memory attached what the team learned ONE COPY — YOUR COMPUTER Files change on your own disk Snapshot saved git commit A SECOND COPY EXISTS Snapshot sent away git push Site rebuilds live for everyone Only the green arrow changes anything about safety — it is the moment the work leaves your computer. A saved snapshot is not a backup: it sits on the same disk as the files it describes.
Where your work lives