The library of NeuralOS
Step-by-step guides, copy-paste prompts and open-source projects. Learn from what the community has built — and build what comes next.
26 resources
Shield your agent against prompt injection with NeMo Guardrails
Your AI app no longer just generates text: it reads emails, runs tools, touches money. That means an attacker no longer attacks your code — they talk to your model. NVIDIA's NeMo Guardrails puts programmable rails between the user and the LLM: it filters what comes IN (jailbreak, injection) and moderates what goes OUT (hallucination, secrets). Let me walk you through putting the first rail in place without losing your mind.
Series · 25/26GuideAdvancedStop guessing whether your prompt works · evaluate agents and skills with Promptfoo
Your prompt "seems to work" — but that's an impression, not data. Promptfoo (the evals framework used by OpenAI and Anthropic) puts a number on quality: you write test cases, compare Claude/GPT/Gemini side by side, measure task+cost+latency, and run it every night in CI. Free, MIT, on GitHub.
Series · 24/26GuideAdvancedBuild your own MCP server in TypeScript with FastMCP
Go from consuming other people's MCPs to publishing your own. With FastMCP you stand up a server that Claude actually calls — from an 'add' tool in 15 minutes to tools with Zod, resources, prompts, OAuth, and edge deploy.
Series · 23/26GuideAdvancedClaude Code hooks · automate your workflow with the agent's lifecycle events
You've spent weeks telling the AI the same thing: "format it", "don't touch the .env", "run the tests". And sometimes it forgets — because an LLM decides all over again on every turn. Hooks are the leap from "I ask it to remember" to "it happens every time, even when nobody remembers": shell commands that Claude Code runs at deterministic points in its lifecycle. Three copy-paste blocks that auto-format, block an rm -rf, and ping you when the AI needs you. It's the step up from "write the contract" to "make the contract run itself".
Series · 22/26GuideIntermediateOne file to govern all your agents · master AGENTS.md
Using more than one AI to build — or planning to switch models to save money? Every new agent starts blind: it doesn't know how your project is tested, which folders are sacred, or how you want your commits. AGENTS.md is a \"README for agents\": a single plain-text file, in your project's root, read today by 25+ tools (Codex, Cursor, Copilot, Gemini CLI, Claude Code…). You write your rules once and they all obey. Here's how to do it right, with a ready-made template and the prompt that generates your perfect AGENTS.md.
Series · 21/26GuideIntermediateWrite the spec, not the code · spec-driven development with GitHub Spec Kit
Vibe coding feels magical on day 1 and maddening by day 20: the AI moves lightning-fast but toward no fixed destination, and every change breaks something far away. The reason is that the AI has no memory of intent — it guesses differently every time. The fix flips the order: first you write the spec (what the app should do), and the agent generates plan → tasks → code from it. With GitHub Spec Kit (122k★, from GitHub, MIT) you install it as a cascade of commands inside your own agent.
Series · 20/26GuideAdvancedAI code surgery · the 7-phase protocol that leaves your app spotless on the inside
AI builds fast, but it leaves the code dirty: orphaned files, duplicated logic, `any` everywhere, animations that melt the GPU. On the outside it looks perfect; on the inside it's rotting. This is the 7-phase protocol an agent applies to your project to clean it up and harden it WITHOUT changing a single pixel or a single behavior — just surgery. It's what an expert debugging surgeon does: goes in, cleans up, stitches up, and your app comes out identical on the outside and spotless on the inside.
Series · 19/26GuideIntermediateThe Unbreakable Guardian · the CI that reviews every change before it reaches your users
You build with AI, everything "works" on your screen, you ship the change… and without meaning to you slipped in a secret key, or opened a hole in your database, or broke something that used to work. Here you'll set up an automatic guardian: 7 inspectors that check EVERY change before it reaches your users, and if ONE of them fails, they block the change cold. It costs ~$0, it's copied into 4 files, and it gives you the language of a serious company. Comes with the verified repo and a single master prompt so your AI installs it for you.
Series · 18/26RepoIntermediateWatch: the robotic user that tests your app and warns you before your customers do
Your tests passed, the deploy went green… and login is still broken in production. Watch is a robotic user that walks through your critical flows in a real browser and warns you the moment something breaks or slows down — with the step, the screenshot, and the fix. Free, on GitHub.
Series · 17/26RepoIntermediateSentinel: the guardian that reviews what your AI wrote (and brings you the fix)
Your AI builds fast, but it writes in "optimistic mode" and can't see its own holes. Sentinel hunts them down and, unlike any scanner, it doesn't just tell you what's wrong: it brings you the fix WITH its test. Free, on GitHub.
Series · 16/26GuideAdvancedLoop engineering · stop hand-flying the AI and let it work toward the goal on its own
You've spent weeks feeding the AI instructions one by one: "do this," you wait, you review, "now do this." You're the engine — and when you get tired, everything stops. Loop engineering is the final leap: instead of chatting step by step, you set up a loop where the AI acts, looks at the result, reasons and repeats, on its own, until the goal is met. It's what ties together memory, RAG, skills, C-A-R and agent-browser into a system that moves forward without you.
Series · 15/26GuideAdvancedGive Your AI Eyes · Stop It Building Blind (agent-browser)
Your AI builds a screen, tells you "done"… but it never opened it to check whether it actually works. It builds blind. agent-browser gives it EYES: it opens your app, looks, clicks, checks, and tells you what's broken. And along the way, it browses and automates the web for you. Open-source, from Vercel.
Series · 14/26GuideIntermediateBuild your design system BEFORE you build · so your app doesn't look \"randomly AI-made\"
Does your app look like it was made by five different people? Every screen with different colors, different buttons, different spacing. The problem isn't the AI — it's that you asked it for screens before a design system. The FIRST thing you should ask it for is exactly that, with a prompt that gives it the level of a Silicon Valley product.
Series · 13/26GuideIntermediateClaude in VS Code · put 3 AIs to work in parallel and move three times as fast
Is your project big and you want to move faster? Instead of a single AI doing one thing at a time, you put SEVERAL sessions working in parallel — each in its own isolated version — and you advance on three fronts at once. Without them stepping on each other, without breaking the good version. This is how we work. They're called worktrees, and with Claude in VS Code it's a single command.
Series · 12/26GuideIntermediateBuild your own skills · teach your AI the way you work (once and for all)
Tired of pasting the same mega-prompt into your AI in every chat? Skill Creator is Anthropic's official tool for building your own "skills" — you teach it your process ONCE, and it's there forever. And the best part: you do it by talking, no code required. Here's how, with the prompt ready to go.
Series · 11/26GuideIntermediateDelete the account without Apple killing your app · rule 5.1.1 done right
Apple rejects your app if it lets users create an account but not delete it FROM WITHIN — and deactivating doesn't count. The advice going around social media (hiding the data) violates GDPR and gets you rejected anyway. Here's the only correct way to do it, with a ready-to-use prompt and how not to lose your business in the process.
Series · 10/26GuideAdvancedEnterprise-grade hardening · the 8 layers that make your app "the hard house"
Nothing is impenetrable — not Google, not the banks — and anyone who tells you otherwise is lying. The goal isn't to be invulnerable: it's to be so expensive to attack that the thief moves on to the neighbor. These are the 8 "defense in depth" layers serious companies use, translated into plain language, with a prompt to ask your AI for each one.
Series · 9/26GuideIntermediateLock down your app before you ship it · the 3 locks AI usually forgets
Building with AI is magical until you push the app to the internet with NO security and anyone can see your users' data just by changing a number in the URL. Here are the 3 most-forgotten locks (RLS, CORS, headers), with simple analogies and ready-to-use prompts — including the trick that keeps your security from making your app painfully slow.
Series · 8/26GuideIntermediateThe C-A-R protocol · how to build software without bugs using your AI (Claude Code, Codex, NeuralOS)
The method that separates building from auditing so your AI doesn't hand you code with hidden bugs. Includes the complete master prompt, ready to paste.
Series · 7/26GuideIntermediateGraphify · give your AI the full map of your project (the most advanced piece)
Memory remembers, RAG searches by similarity… but neither one understands how your whole project CONNECTS together. When your AI starts getting lost, breaking things, and burning tokens reading blind, you give it a graph: a map of what depends on what. Open-source, free, a skill you install in one command.
Series · 6/26GuideAdvancedThe "brain" · give your app a mind of its own (not an expensive parrot)
An AI app that forgets every user and knows nothing about your business is an expensive parrot. A "brain" turns it into an employee who knows your operation and remembers every customer. It's memory + RAG + graph, together, at the product level. Here's what it is, when you truly need it, and how to build it without over-engineering.
Series · 5/26GuideIntermediateRAG without the hype · when you ACTUALLY need it (and when it's throwing money away)
RAG used well saves you a TON of tokens (and money) when your information is large — but Anthropic themselves say that if it fits in ~500 pages, setting it up is throwing money away. Here's the rule for knowing when you SHOULD, and the THREE ways to implement it, from the simplest (no code) to the one behind a super-app — including how ours works.
Series · 4/26GuideBeginnerBack everything up to GitHub before AI breaks your project · your infinite undo button
The AI touched 40 files and now nothing opens. You don't always lose everything… but one wrong command or an accidental delete really can leave you with no way back. GitHub is your own safety net: you return to the exact moment it was working, with one click. And most importantly: when and how to use it every single day.
Series · 3/26GuideIntermediateKarpathy's "LLM Wiki" method · make your AI actually accumulate knowledge
Andrej Karpathy (co-founder of OpenAI) published a pattern that went around the world: instead of explaining everything to your AI over and over, have it build and maintain its own wiki — a knowledge base that grows on its own. Here's the full method, in plain terms, with the prompt to set it up.
Series · 2/26GuideBeginnerYour AI's memory isn't controlled by you · the tools that put it on your side
Yes, ChatGPT and Claude already \"remember\" some things between chats. The problem: you don't know what they store, you can't read it or carry it to another model. Here's the ladder of REAL tools to give your AI a memory that IS yours, from the easiest to the most powerful.
Start hereGuideBeginnerTurn your AI into your office team · 5 official Anthropic skills (with the command for each one)
Your AI can build the PowerPoint, clean up the Excel, fill out a PDF and review your app — all on its own. These are 5 official Anthropic skills, and they install with a single command. With the command and a ready-to-use prompt for each one.
We'll let you know when your wave opens.
And the most important AI news of the week, without the hype.