A candid 50-minute walkthrough of the tools, techniques, and mindset shifts that have transformed the way I build — shared with friends over a morning call.
Using "ask me questions" to let AI build the thinking framework for you — not the other way around.
Persistent knowledge bases attached to chat sessions — your context, always there.
Giving Claude memory between sessions by writing decisions and state to markdown files.
Retrieval-augmented generation — an index that lets AI search massive docs without hitting context limits.
Click each topic to explore the technique, the reasoning, and the practical tips.
In the old days you'd design a full framework first — titles, subheadings, structure — and then fill in content. AI inverts this. You start with a half-formed idea and use AI to rapidly build the framework for you.
The key insight isn't just "write a good prompt." It's telling Claude: "ask me questions to clarify." That single phrase offloads all the meta-thinking onto the model. It figures out what's missing. You just answer.
"I didn't have to think what publication it's for, what the tone is, or any of this. I could just start with: write me an article about rabbits." — Adriaan
E.g. "I want to write an article about rabbits." Don't over-engineer it — let AI do that.
Word count, target reader, publication, tone — whatever constraints you know. Even just one or two helps.
Claude Opus will respond with multiple-choice options. You pick. No blank-page paralysis.
Claude builds on each answer. The output framework gets richer with every loop — minimal effort from you.
The new chat gets a fully detailed brief. Claude can now web-search, fill in data, and write the full piece.
A Project in Claude is a persistent knowledge base attached to every conversation in it. Drag documents in once — they're referenced forever. No more copy-pasting background context into every new chat.
Free tier gives you 5 projects. Paid gives more. Each project can also hold instructions (run on every new conversation) and memories (short summaries Claude maintains for you).
"Instead of having to save it as a text file somewhere and copy paste for every new conversation, all these conversations are referencing this kind of knowledge." — Adriaan
Adriaan dragged all 6 books (free text files) into a project, then used Claude to build a playable game based on the story world.
Pasted all the system context, pilot results, and background once. Every conversation about the hospital pitch has full context — no re-pasting.
Fed in weekly payout data from payment aggregator. Claude built interactive dashboards showing top products, revenue breakdowns, and month-on-month trends.
Point Claude at a website you admire. Ask it to extract all CSS variables, font choices, colour logic, and spacing into a "style guide" doc — then use that doc across all future projects.
Claude Code is the terminal-based version of Claude that has direct access to files on your computer. You give it a specific folder — and only that folder. It can read, write, and edit files inside that folder.
The big problem: Claude has no memory between sessions. Every new terminal session, it starts fresh. The solution is a Design Log — markdown files Claude writes to as it works, so it can always pick up where it left off.
"By writing it down you give it memory — it sees: this is what I was doing, this is what I've done, this is what still has to be done." — Adriaan
This file is read automatically every time Claude Code starts. It contains: who the project is for, what tech is used, allowed tools, things Claude must NOT do.
A full design document (can be 500+ lines). Covers architecture, goals, decisions made, and the "why" behind everything. Claude references this when it needs context.
Claude writes its own tasks here as it works. Each task has status. If a session crashes or times out, the next session reads task.md and continues seamlessly.
Before executing, Claude enters a planning mode: it asks questions, writes a task list, and only then starts implementing. Think "measure twice, cut once" — but for AI.
Every AI model has a context window — a limit on how many words it can hold in "mind" at once. Think of it like a desk. Only so many pages fit on it. When your document library is huge (100,000+ words), you can't put all of it on the desk.
RAG solves this by adding an index — like a book's table of contents. Instead of reading the whole library, Claude reads the index first, finds which sections are relevant, and only pulls those sections onto the desk.
"Instead of having 100,000 words, now it only needs to look at 2,000 — the relevant documents." — Adriaan
Claude can hold ~200,000 tokens at once. Sounds like a lot — until you have years of meeting notes, a full codebase, or an entire book series.
Tell Claude to scan all your documents and create a table of contents: document name, date, key topics, page numbers. This becomes a tiny "map" of your knowledge base.
Future queries say: "use the index to find relevant files, then read only those." Claude goes from reading 100k words to reading 2k — much faster, much cheaper.
Enterprise RAG uses semantic search databases (Pinecone, Weaviate, pgvector) that store document "embeddings" — mathematical representations of meaning. But for home use, a simple index file works great.
Every technique above solves the same problem: AI is only as good as the context you give it. Prompt expansion builds context up-front. Projects preserve context between sessions. Design logs rebuild context after a crash. RAG retrieves context from large stores efficiently. Master context management and you master AI.
Every one of these was designed and built using only Claude Code, Claude, and Perplexity. No manual coding.
Home management app for Rosemary. Rooms, tasks, shopping lists, recipes with ingredients & method, notes, appointments — all in one place. Built because 5 different apps is too many.
Passkey AuthFactory floor trolley tracking. Operators scan or tap through stations (cutting → bending → grinding). Real-time visibility on where each trolley is in the production flow.
OperationsPersonal finance dashboard. Drag and drop a bank statement — it parses transactions and lets you categorise each as expense, tax payment, or transfer. Replaces spreadsheets.
FinanceRFID-based point-of-sale for the office snack shop. Tap your badge → select items → logged instantly. Replaces the paper book people used to write in when they took a snack.
POSPersonal blog running from the same single server. Features ginger cookie recipes, engineering notes, and reflections. Because why not host your thoughts next to your factory app.
ContentA Windows game recreation of an old Nintendo DS title. Full game design document (GDD), phased implementation plan, task tracking — all managed via Claude Code and design log files.
Game DevThis very page demonstrates everything discussed in the session. The transcript was pasted into Claude Code, which wrote all the HTML — then it was shipped live using SSH, Docker, and Nginx.
The container uses nginx:alpine — a tiny, secure Linux image (~5MB).
It joins the proxy-net Docker network so Nginx Proxy Manager can reach it without exposing ports directly to the internet.
Port 8087 is the external port Nginx Proxy Manager forwards to.
Inside the container, Nginx serves on port 80 as normal.
The result: a single static HTML file, served globally, from a €3/month VPS in Germany.
Key moments from the 50-minute session with Adriaan, Marni, Josh, Renske, and others.