Andrej Karpathy 亲述：99% 的 AI 用户不知道的 7 个基本功

Picture this.

It's 11pm. You've been staring at the same AI chat window for two hours. You've rephrased the same request six different ways. You've tried being polite, being blunt, being specific, being vague. You've copy-pasted three different "magic prompts" from some guy on X who swears his template makes Claude "10x smarter."

Nothing works the way you expected. The output is either too generic, structurally wrong, or confidently incorrect about something you already told it twenty minutes ago – in this very same conversation.

You close the tab. You'll try again tomorrow. Maybe a different model. Maybe a different prompt. Maybe you're just not good at this yet.

Here's the uncomfortable truth: it's probably not the model. And it's definitely not the prompt.

While most people are endlessly tweaking their wording, hunting for the perfect instruction, or buying yet another "AI productivity course" – a small group of people quietly figured out that the problem was never the prompt at all.

The problem was everything around the prompt.

The context. The memory. The structure. The workflow.

Andrej Karpathy is one of those people. And unlike most voices in the AI space, he has receipts: co-founder of OpenAI, former head of AI at Tesla, Stanford professor, one of the engineers who actually built the systems everyone else is trying to "hack" with clever prompting.

He's been thinking about this longer than almost anyone. And what he's concluded is both obvious in hindsight and almost completely ignored in practice.

He doesn't use magic prompts. He builds infrastructure.

Seven habits. A few simple files. A specific working rhythm. That's it.

Here's exactly what he does – and why each piece matters.

TIP 1: Forget magic prompts. The problem is almost always missing context.

Since 2022, "prompt engineering gurus" have dominated X and Instagram.

The message: learn the right spell and the model obeys.

Karpathy disagrees. The real reason most people iterate 100 times and still get bad output? They ignore context entirely.

His actual formula:

• Write a standard, clear request
• Always include a concrete example of what good output looks like
• Paste the full error message or full background – never a trimmed snippet

Don't cut your code or text to "save context window." When the model guesses what's missing, it gets it wrong. Every time.

No secret instructions will teleport your background into the model's head. You have to write it out.

TIP 2: Your CLAUDE.md is probably garbage. Go check it right now.

Did you copy it from someone else's template? Did you let Claude write it for itself? Then that file is not working for you.

Your main config file must clearly explain five things:

• Who you are
• What the project is (general frame only)
• What not to touch
• File naming conventions
• How to format responses Almost everyone has the file. Almost nobody has set it up properly.

Before you blame the model for being "dumb" – go read your own instructions to it.

And if you only use browser-based AI tools? You still need this. Set a pinned brief. Same logic applies.

TIP 3: Build a three-layer system. Stop restarting from zero every session.

Karpathy's pipeline:

• /raw – your raw source material, dumped in as-is
• /wiki – structured pages the model writes and maintains
• CLAUDE.md – your standing operating principles New source comes in → drop it in /raw → tell the model to process it.

That's 30 minutes saved per day, compounding.

If your project lives longer than a couple of days and you're re-explaining everything in every new session – that's not a workflow, that's a loop.

TIP 4: After every strong answer – save it. Permanently.

The default habit: get a great response, copy the result, close the tab, forget it. Karpathy says this is quietly killing your long-term productivity. Models need references.

After every valuable response:

"Save this as a permanent page: wiki/topic/.md"

Then audit your notes periodically for duplicates, conflicts, and outdated information.

Skip this and your best AI outputs quietly drown in chat history. You'll spend hours on tasks you already solved.

TIP 5: For any project lasting over a week – add index.md and log.md. No exceptions.

Two files. Two purposes:

• index.md – a map of everything that exists
• log.md – a running changelog: date | type | description

Example: 28-05-2026 | summary | customer interview breakdown

If you vibe-code 1–2 hours a day, in two weeks you genuinely won't remember what you built on day three. These two files are your memory layer.

TIP 6: AI is a brilliant intern with no taste. Treat it like one.

Karpathy's framing: AI agents are "superpowered interns with massive knowledge, who constantly hallucinate and have zero taste for code." They need a tight leash.

His actual working loop:

• Load full context
• Ask for 2–3 options for the next small step only
• Pick one
• Evaluate, test, commit
• Repeat

Never ask it to do everything in one prompt. That's how you get 500 lines of undebuggable mess.

TIP 7: One sentence that makes every research prompt 10x more readable.

Add this to the end of any analysis or research prompt:

"Structure your final response as a self-contained HTML file."

AI models render anything into clean, navigable HTML in seconds. Reading time drops dramatically. It costs you one sentence. Use it every time.

Here's what's strange about all of this.

None of these tips are secret. None of them require a paid subscription, a special tool, or a 40-hour course. They're all, once you see them, completely obvious. Of course the model needs full context. Of course you should save what works. Of course a project needs a map and a log.

And yet – go look at how you actually use AI right now. Be honest. How many of these seven things are actually in place in your workflow today?

Most people are in a strange place with AI. They believe it's powerful – they've seen it do impressive things – but in their own hands it keeps underperforming. So they assume the gap is about the model, or the prompt, or some insider knowledge they haven't found yet. They spend hours searching for the trick instead of spending twenty minutes building the foundation.

Karpathy's entire message is that the gap isn't about magic. It's about memory, structure, and incrementalism. Give the model your full picture. Save what it builds. Work in small, committed steps. The model isn't the bottleneck – your workflow is.

The people who will get dramatically more out of AI over the next two years aren't the ones who found the best prompts. They're the ones who built the best systems around the model – even simple ones. A /raw folder, a /wiki, a proper CLAUDE.md, two markdown files, and a working loop.

That's the whole edge. It's almost embarrassingly small. But almost nobody is doing it.

Go back to the story at the top. That person at 11pm, frustrated, closing the tab – that's not a story about a bad AI. That's a story about a workflow with no memory, no structure, and no incremental loop. The model was ready to help. It just didn't know enough about what it was helping with.

Now you know what to build. Start with one file. One folder. One saved response. The system compounds fast.

TL;DR

Stop tweaking prompts. Start building infrastructure. A proper config file, a /raw and /wiki structure, permanent reference pages, index and log files for long projects, a small-steps working loop, and one HTML trick. The model stops guessing – and starts actually helping. The edge isn't a secret. It's a system. And it takes about an afternoon to set up.

If this was useful – bookmark it. You'll want to come back to it.

Follow @ScottyBeamIO for more breakdowns like this.

No fluff, just what actually works.