every project on this site was built without a single line of engineering experience. here is how, and how you can start.
scroll to watch it builda real prompt on the left. a real result on the right. then it rebuilds the very page you are on.
Nope. Not one line. Okay… maybe a couple of lines 😉 Every project on this site was built through conversation with AI. The interactive essay on how AI thinks. The meditative snake game. The cinematic chapters at atom738. The bible study. The AI OS field guide. All of it.
The loop is simple. You describe what you want. The AI writes it. You look at the result, ask for changes, and keep going until it works. There is no point where you need to read code you do not understand.
The skill is not writing software. The skill is knowing what you want and being able to describe it clearly. If you can write a decent email, you can do this.
not screenshots. live pages, same backgrounds and animations. click to open, scroll inside.
Most AI forgets you the moment the tab closes. Point it at your Obsidian vault — a plain folder of markdown — and it gains a memory that stays. Here's the whole journey, one chapter at a time.
They sit in a folder, slowly decaying — and every chat AI forgets you the second the tab closes. Point an agent at that folder and the notes wake up: it reads them, files new ones, writes back to them, and keeps working while you sleep. Everything ahead is that one shift, unpacked.
No database, no special format — a vault is just files on disk. Six steps take you from a pile of notes to a living system the agent can read, write to, and grow.
Put your notes in one Obsidian vault — a plain folder of .md files — and make a copy, or put it under Git, before any AI touches it.
in plain words a "vault" is just a folder on your computer holding your notes as simple text files you can open in anything. Before you let an AI rearrange them, make a second copy — drag the folder to a backup drive, or use a free tool called Git that keeps a timeline of every change so you can undo to any earlier version.
why the files are the asset; the AI is disposable. A backup means a bad bulk edit is one revert away, not a loss.
Run Claude Code inside the vault directory for the agentic setup, or add a filesystem MCP server (like obsidian-mcp-server) pointed at your vault path in Claude Desktop.
in plain words the AI needs your permission to open that folder. Two easy ways: run a free app called Claude Code from inside the folder, or switch on a small "connector" (that's what MCP means) in the Claude desktop app and tell it where your folder is. It's a one-time setup, and you're not writing code.
why a vault is just files on disk, so giving an AI read/write access is one command or one block of config — no special format, no database.
Add a CLAUDE.md at the vault root: who you are, your folder conventions, and a session protocol — read the inbox at the start, write a dated log at the end.
in plain words AIs forget everything the moment you close the chat. So you leave a note-to-self file (named CLAUDE.md) at the top of your folder: here's who I am, here's how my folders are organized, do this at the start and end of each session. It reads that first every time, so it already knows how you work.
why continuity lives in files, not the model. The agent reads this every session, so it resumes instead of starting cold.
A few folders — Inbox, Projects, Resources, Archive, AI/sessions — and YAML frontmatter on notes (created, tags, status, related). Cap it at about three tags a note.
in plain words don't bury notes in folders inside folders inside folders — keep a handful of top-level ones. "Frontmatter" is just a few label lines at the very top of a note (the date, a tag or two, a status), like a sticker on a file folder, so the AI can sort and find things fast instead of reading every note.
why manual filing breaks past ~1,000 notes. Flat folders plus consistent frontmatter let the agent filter and cross-reference instead of hunting.
Install Smart Connections or Copilot inside Obsidian for semantic search and related-note surfacing; point them at a local model via Ollama if you want everything on-device.
in plain words add a plugin (Smart Connections or Copilot) so you can chat with your own notes and get related ones suggested automatically. "By meaning" means it can find your note about "dad's watch shop" even when you search "father's business." Want full privacy? A free tool called Ollama runs the AI on your own computer, so nothing ever leaves it.
why embeddings find notes by meaning, not keywords — and the local option keeps private notes off any cloud while still giving you chat-with-your-notes.
Commit the vault and the agent's memory files regularly, review what the AI wrote back before trusting it, and revert when a session goes off the rails.
in plain words make it a habit, and keep that Git "undo timeline" running ("commit" just means saving a snapshot you can return to). Skim what the AI changed before you trust it — and if a session makes a mess, you roll back to yesterday's version in seconds. Nothing is permanent.
why value compounds only if you audit it. Git makes the knowledge revertable and inspectable, so one bad session can't quietly poison the rest.
Once it's running, the agent cycles the same four moves — and each pass leaves the next one smarter.
Dump a raw paragraph, the text off a screenshot, or a rambling five-minute voice memo. It works out what the thing is actually about, then writes a clean note — titled, tagged in frontmatter, dropped in the right folder, and wikilinked to what it relates to. You capture at the speed of thought and never file by hand again.
under the hood it writes plain .md straight to your disk — nothing is trapped inside an app.
Ask "what did I decide about the move?" and the answer comes from your vault, not the open web — with a link to the exact page it came from. This is retrieval aimed at you: every claim traces back to something you actually wrote, so it describes your life without inventing it.
under the hood it searches your notes by meaning and by their link graph, then quotes the source.
Turn it loose on a vault you abandoned years ago. It reads each note as an idea, proposes a real structure — folders, maps of content, a consistent tag system — then performs the moves and rewrites one approved step at a time. Nothing is lost, because every change is reversible text.
under the hood it proposes, you approve, it executes — and version history remembers every step.
It keeps finding links you never drew yourself, threading notes written years apart. An idea you gave up on resurfaces the moment a new one makes it useful. Your thinking stops decaying in a drawer and starts compounding.
under the hood every new note re-threads the graph of connections around it.
Now the busywork disappears and the vault becomes an extension of how you think — draft a decision from your own past reasoning, plan your week from one plain-text file, wake up to a morning brief pulled from notes you forgot you made. A few to start with:
Point the agent at years of messy clippings and let it synthesize a clean, cross-linked wiki of entity and topic pages. The work happens once at ingestion, then compounds.
When the AI answers a real question from your notes with citations, save that answer as its own note. Your explorations stop evaporating and join the searchable base.
Write freely in a daily note, then have the agent reflect using what it already knows about you — citing the specific past notes it drew on. Your own history becomes the context.
Describe an automation in plain language — a tracker, a restructuring script, a tiny app — and let the agent generate and wire it. The skill is articulating the idea, not the syntax.
Have it rewrite inconsistent frontmatter across thousands of notes, rename cryptic folders, or merge scattered dailies into themed logs. Work never worth doing by hand becomes a ten-minute job.
Set the agent to read your inbox and today's note at the start and write a dated log at the end. Each session begins smarter because the last one wrote down what it learned.
Not theory — real vaults, real tools, real numbers.
Writer Eleanor Konik connected Claude to a vault far too large for normal search via a filesystem MCP server — semantic answers in seconds, years of notes batch-restructured with AI-generated scripts, frontmatter standardized across the whole thing in under ten minutes.
In Andrej Karpathy's pattern, the agent reads each new source and integrates it into a living markdown wiki, updating entity and topic pages. One source can touch ten to fifteen pages — the wiki becomes a textbook you wrote, not an index you re-search.
A reporter handed a disorganized vault to Claude with a six-phase prompt: ZIP backup, then PARA folders, capped-tag frontmatter, wikilinks proposed for approval before being added, and a generated index. A repeatable recipe — with the honest caveat that moving files can break wikilinks.
A growth marketer built a clothing-recommendation app that reached 85,000 users; a parent built a carb-counting app for their kid's diabetes; a design lead chained tools into an agent that writes in his voice — no hand-written code. The shared report: the hard part is a good conversation about what you want.
An analysis of one multi-year vault (8,053 notes, 64,890 links) found link density rising with synthesis: daily notes averaged about 11 links, yearly summaries over 1,400. The value lived in the connections, not the raw collection.
Strip away the demos and here's what's actually left.
Plain markdown in a folder you own can be read on a 1960s machine and a 2160s one. Apps die and companies fold; a local text file you control survives the churn and ports to whatever comes next.
Everything is a file you can open, so you see exactly what the AI sees. Memory is inspectable and editable — not an opaque cloud store where you're a borrower of your own data.
Each conversation can write a note back, and each future session reads it, so the base gets smarter over time. Synthesis done once at ingestion beats re-piecing fragments on every query.
You no longer need to know how to code to build and reorganize systems. You need to imagine clearly, describe precisely, and judge what comes back — which genuinely opens building to everyone.
It only pays off if you synthesize instead of hoard, keep the structure simple, and version your memory. An unaudited, append-only memory file lets one bad session quietly poison every future answer — so commit, and review.
sources Karpathy · LLM wiki·Konik · Claude + Obsidian MCP·HowToGeek · vault cleanup·Ango · file over app·Smart Connections·PKM at scale · 8k notes·Harvard · vibe coding
One of these is not like the others. Obsidian is home base, and the only piece that is not AI. The other three are AI apps that all work inside it. You do not need them all. Tap any card.
obsidian not AIhome base · where everything lives
A free notes app, and the one tool here that is not AI. Everything you make lives inside it as plain text, on your own computer. It remembers everything, locally and private, so nothing ever leaves your machine.
how it works. You organize your whole life here by letting AI work directly inside this vault. Obsidian is the home; the three AI apps below all plug into it and read and write these same files. You can switch between them freely, your vault never changes.
obsidian.mdthe three AI apps · all work inside your vault
kiro my favoritethe AI builder I reach for first
An AI builder you talk to. Describe what you want and watch it write your project's files in real time, while you steer in plain language the whole way.
why it is my favorite. It is the one I reach for first. If you have never built a website or an app, this is where it appears, piece by piece, without you writing a single line.
kiro.dev
claude coworkthe coworker that runs on its own
Claude working on its own. You hand it a whole task, plan a series, organize the vault, research a topic, draft a chapter, and it carries it out across your files by itself, checking with you before anything big.
cowork vs claude code. Cowork is hands-off: you give it a goal and it runs, on its own, in the background. Claude Code (below) is hands-on: you direct it move by move. Reach for Cowork to think and organize your life, and Claude Code to build.
claude.com/product/cowork
claude codethe hands-on builder you direct
Claude as a precise builder you steer step by step. Where Cowork runs on its own, Claude Code does exactly what you direct, one move at a time, which makes it the best of the three for making real things.
I run it inside Visual Studio Code, a free code editor. That is my favorite way to work with Claude Code. There is a longer write-up in the workshop.
All four have a real free or low-cost tier, and none require any coding background. Here is what each one gives you, and what I actually run.
Free, forever. The whole app: your notes as plain text on your own computer, unlimited, no account, nothing sent to a server. Two optional add-ons if you want them, Obsidian Sync (encrypted, across your devices) and Obsidian Publish (turn a vault into a website).
A free tier to try it. Claude Pro unlocks both Cowork and Claude Code. Claude Max is the same, with far more daily usage (its 5x and 20x tiers), which is what you want when you are building for hours. Max is the membership I run.
A real free tier with working models. The paid tiers (Pro, Pro+, Power) add more room to build and premium models like Claude Opus and Sonnet. I keep a paid plan, since this is where most of my building happens.
Always free. It is just the editor I open Claude Code inside. Nothing to buy.
Want the exact stack behind this site? Obsidian (free) + Claude Max + a paid Kiro plan, all driven from VS Code. Start free, and only upgrade the one you actually hit limits on.
It starts in Claude Cowork. You point it at your Obsidian vault, the folder where all your thinking lives, and from then on Claude works inside it with you. It reads your notes, writes new ones, organizes the mess, and holds the thread across days.
So a build begins as a conversation. You talk through what you are making: the problem, the audience, the feeling. Claude takes it down as you go, drops it into the vault, and links it to what is already there. By the time you stop talking, the plan exists as real files you can see.
Then the building. You hand that plan to kiro or claude code and describe the first piece. The AI builds it. You look. If it works, ask for the next thing. If it is wrong, say what is wrong. If it is ugly, say so.
The whole time, Cowork is still holding the context, so you never start from scratch and never re-explain yourself. You stay in your own vocabulary, from the first sentence to the finished thing.
the whole session is one loop
tap a stage ↑
Open a chat and describe the thing in plain language — the problem, who it is for, the feeling you want it to have. No syntax, no jargon. The clearer the picture you paint, the better the first draft comes back. This is the only part that is truly on you.
Claude turns your description into the real work — the files, the code, the words, the layout — and shows it to you as it goes. What used to take a team and weeks arrives in seconds, as something you can actually look at and judge.
Look at what came back and respond like a director: make the header bigger, this tone is off, now add a contact form. You are not editing code — you are giving notes. Each reaction sharpens the next version, and the loop runs again.
You stop when it is right. That is the whole skill.
One markdown file with everything you need to begin.
Inside: a folder structure that works on day one, a starter instructions file (the document that tells the AI how you like to work, how you think, what your projects are), and a short walkthrough to get the four tools talking to each other.
Download it. Open it in obsidian. You have a workspace.
download the starter kitCreate a new folder on your computer and move starter-kit.md inside. That folder is your workspace.
Get it at obsidian.md, choose "open folder as vault," and pick that folder.
Open Claude Cowork (or Kiro) and give it access to that folder, so the AI can read and write your files.
Open the ABOUT-ME file and answer a few questions about how you like to work. This is what makes the AI feel like it already knows you.
Tell it what you want to make, in plain words. Watch it build, and keep going until it is right.
If you make something with it, send it. I want to see what you build.
