All posts

· 6 min read

You don't need prompt engineering: just ask

Every week there’s a new course, a new framework or a thread with “the 37 secret prompts” for coding with AI. My take is a lot less exciting: coding LLMs are already good enough for you to simply ask. Say clearly what you want, discuss, test, talk it through. That’s it.

I’m not alone here. The Just Prompt, Motherfucker manifesto, written by Carlos Eduardo Coelho and Leandro Proença, says the same thing with far less politeness, and I agree with almost all of it. Below is why, with sources.

Prompt engineering solved a problem that barely exists anymore

“Think step by step”, “you are a senior engineer with 20 years of experience”, promising a tip, threatening the model, writing CRITICAL RULE in all caps. All of that started as a workaround for weak models. In 2023 it even made a difference. Today, when someone actually measures it, the effect disappears:

  • Wharton’s Prompting Science Report 2 tested the famous “think step by step”. On models that already reason on their own, the gain was marginal and responses took 20% to 80% longer.
  • Prompting Science Report 3, from the same group, tested threatening the model and offering it a tip. The conclusion is that this generally has no significant effect on performance.

And some of it went from useless to harmful. Anthropic’s prompting documentation warns that newer models follow what’s written more closely, and that the aggressive language of the old days now makes the model overdo it. Their recommendation is literally to tone it down: where you used to write “CRITICAL: You MUST use this tool when…”, write “Use this tool when…”.

In other words, the shouty prompt you copied from a 2024 repository may be getting in your way.

Anthropic and OpenAI already optimized the product for you

When you open Claude Code, Codex or Cursor, you’re not talking to a raw model. There’s a system prompt, a set of tools and context management put together by the people who trained the model and measure the outcome at scale. That part of the “engineering” is already done, and by people with far more data than the guy selling a course.

What’s left for you is the human part. The companies say so themselves:

  • Anthropic’s golden rule is to show your prompt to a colleague with little context on the task. If they’d be confused, the model will be too.
  • OpenAI’s prompting guide for GPT-5.6 says to describe the destination instead of prescribing every step. In their internal runs with coding agents, leaner system prompts improved scores by roughly 10% to 15% and cut 41% to 66% of the tokens. They note that the numbers vary by workload, but the direction is clear: fewer instructions, better results.

Notice that neither of them talks about a secret technique. They talk about clarity.

More tooling means more tokens, and not always better results

This is where people get lost. They install twenty MCP servers, fifteen skills, an agent orchestrator and an 800-line instructions file, and think they’ve built a machine. What they’ve built is a clogged context.

  • Huge context files. A study evaluated AGENTS.md files with coding agents. They did not generally improve task success rates and raised inference cost by more than 20% on average. According to the authors, agents follow the instructions in the file well, but that big repository overview doesn’t help. What’s worth keeping is a record of your project’s non-standard practices.
  • Too many MCP servers. Anthropic itself, which created the protocol, explains that most clients load every tool definition into the context before reading your request, which makes the agent slower and more expensive.
  • A full context makes the model worse. Chroma’s Context Rot report tested 18 models and showed that performance doesn’t hold up as the input grows, even on simple tasks.

The manifesto sums this up in a table I find perfect: “Twenty MCP servers” on one side, “Nineteen you never call” on the other. And all of them taking up room in the conversation.

How to code with AI without advanced techniques

What works is what would work with a good programmer who just joined the project:

  1. Say what you want and why. The goal, what can’t break and how you’ll know it’s done.
  2. Discuss before asking for code. Ask for a plan, question it, disagree. It’s cheap and saves rework.
  3. Test. Run the code. If it errors, paste the error. If it looks bad, say it looks bad.
  4. Correct course by talking. You don’t need to nail the prompt on the first try. You need to reply.
  5. Keep the fixed instructions short. Only what the model can’t figure out by reading the repository: the test command, a team convention, what’s off limits.

A good request looks like this:

Checkout is charging shipping twice when the customer changes the address
after applying a coupon. Find the cause before touching anything and explain
it to me. Then fix it and write a test that would fail without the fix.
Don't touch the carrier integration.

No persona, no all caps, no template. Just a well-explained problem.

So who’s selling the advanced technique?

In my opinion, most of the people selling advanced prompt engineering, miracle skills and methods with fancy names are just hustlers who left the house looking for suckers. The product isn’t your working code. The product is the course, the subscription, the harness, and your feeling that you’re falling behind.

This isn’t a grudge against tools. The manifesto’s own authors write “We have nothing against the tools. We use them”. An MCP server you use every day is worth it. So is a skill that captures a tedious process your team has. The problem is piling things up out of fear, without measuring whether anything got better.

The criterion I use comes from Anthropic’s Building Effective Agents: find the simplest solution possible and only add complexity when it demonstrably improves outcomes. If you can’t say what a tool improved, it’s just burning tokens.

To sum up

  • Prompt tricks were a crutch for old models. On current ones they measure as useless or get in the way.
  • Anthropic and OpenAI already optimize the product and ask the same thing of you: clarity.
  • A context clogged with tools and instructions costs more and delivers less.
  • What works is explaining well, discussing, testing and talking it through.

The manifesto closes with a line I’d sign under: “anyone who can explain a problem to a colleague can prompt”. And for those who can’t, no amount of engineering will save them.

If you want to follow the next posts, subscribe to the RSS feed.

Contact

Have a project in mind?

Tell me what you need and I'll get back to you with next steps and an estimate.