Your Old Prompts Are Making Claude Worse: 7 Moments From Anthropic's Talk, and the 6 Commands That Cut the Bill

Most people pay more for Claude and still get worse answers. The reason is usually not the model. It is a handful of small settings: old prompt habits, a broken cache, and the effort dial turned the wrong way.

That is the argument Lance Martin from Anthropic made in a 64-minute talk at AI Engineer, and he showed six commands that fix it. We watched the whole thing and pulled out the seven moments that matter. We also ran the last two commands on a small support-ticket app of our own: accuracy went from 74% to 100% on tickets it had never seen, and the cost per ticket fell from 3.18 cents to 0.33 cents.

1. Your app is not on the frontier

Every AI model sits on a curve. Pay more, get better answers. Martin's point is that your app usually isn't on that curve at all. It sits inside it, so you pay more and get less.

In his words: "Your application is actually sitting inside this frontier that you could be on, and it's because of minor misconfigurations." The rest of the talk is a list of those misconfigurations and how to fix each one.

2. Old prompt habits are now a tax

Think step by step. Double check your work. Be maximally thorough. We all learned those tricks. Martin calls them leftovers from "several generations ago" and says plainly: "You don't need that anymore. Frontier models are very good at thinking."

On today's models those lines are a tax. The model already does all of that, so you pay for the thinking twice. There is a second problem. If two of your rules contradict each other, older models just ignored one. New models get confused.

Some people in the comments on the talk said it's convenient for Anthropic to blame our prompts. Fair. But you don't have to take his word for it, because one command checks it for you:

• /checkup prompt-audit strips old prompt habits from your Claude Code setup.
• /claude-api prompt-audit does the same audit for apps that call the API (the way software talks to Claude directly).

Not a coder? Open your Claude project instructions and delete any line that tells it to think step by step.

3. The prompt cache is free money

When an agent works, it sends the same instructions again on every single turn. If those instructions come from the cache (a saved copy Claude has already read), you pay about a tenth of the price.

The catch is that the cache only works if the start of your prompt is exactly the same each time. Martin: "Prompt cache are a strict prefix match. So even a byte change in your prefix will invalidate the cache." Put today's date or a timestamp at the top, and every turn pays full price.

The fix is /claude-api cost-optimize. It moves the parts that never change to the front and the parts that change to the end. Martin and Anthropic's Applied AI team ran it on a set of internal benchmarks and "found opportunities to reduce costs by up to 72% with no meaningful change in performance."

So the savings ranged from about a quarter to almost three quarters of the bill. If you build on the API, run this one first.

4. Smart model, low effort (the 39:24 moment)

This one is buried 39 minutes in, in the Q&A. Most people watching an hour-long talk never get that far. It's about the effort setting, the dial that controls how hard the model thinks before answering.

Most people pick a cheaper model and crank the effort to max. Martin says to do the opposite: "Moving to higher models at lower effort is often cheaper and more performant than lower models at high effort." And: "Try low to start. You can always dial back up." Max effort, he says, is "extremely token hungry."

In his support-ticket test, Opus 4.8 on high effort scored 74.4%. The tool moved to a stronger model on low effort, then stepped down to Sonnet 5 on low, then tuned the prompt. Accuracy went up, and the cost per ticket fell by about 90%.

On claude.ai the same idea works. Pick the smart model, and leave extended thinking off for everyday questions.

5. Evals Claude can actually climb

How does Claude know it's getting better? You need a test set, called an eval. Martin's rules are simple:

• Grade pass or fail, not one to five. "Grade it from zero to five. It's very ambiguous. Make it crisp and simple."
• Have two different models grade the same answers. If they disagree, your grading rules are broken.
• Remove the questions nobody can answer. Otherwise it looks like there is 20% more to gain, "but no model can actually achieve it."

The command is /claude-api build-eval. Point it at your real logs. It interviews you, flags the tricky cases, and builds the test with you step by step.

We ran this in September on a small support-ticket app. It measured our starting point at 74%, and it told us exactly which line in our prompt was causing the misses.

6. Hill climb overnight (our real run)

Once you have a good test, Claude can improve your prompt while you sleep. The command is /claude-api hillclimb. Martin describes the goal: give an agent an evaluation, let it run, and "come back, for example, in the morning and it's significantly improved."

But there's a trap. He warns that you can spend a lot of tokens, see a great score, "run it in production and performance is not better at all." That's overfitting. Claude learns your test instead of your job.

So the command splits your tests into two piles. Claude studies one pile and never sees the other. If the hidden pile doesn't improve, the change gets thrown out. Here is our run:

About three cents to a third of a cent per ticket. The same 90% cut he showed on stage, on our own app.

7. Agents that work for days

Moment seven is where all of this is heading: agents that work for days. Martin's rule is to keep the brain and the hands apart.

He gives the reason: "Imagine you have an agent that you spin up overnight to just run, and the sandbox dies midway through the night. You lose state, you lose your session log and everything." So the agent's memory lives in a log that survives a crash. If the sandbox (the sealed workspace the agent works in) dies, a new one spins up and picks up where it left off. And your secret keys stay in a vault. The agent can use them, but it can never read them.

To try a working example, run /claude-api managed-agents-onboard.

To be fair

The "old prompts" argument comes from the company that sells the model, and viewers pointed that out. The audit commands let you check it on your own setup instead of trusting the claim.

The effort advice is not universal either. One viewer wrote that for many medium-complexity coding tasks, it was cheaper to run Haiku at max than Sonnet at lower levels. That can happen, because cheaper models have cheaper tokens. So don't guess. Test it on your own task.

The 72% figure is "up to", on Anthropic's own internal benchmarks, and the lowest one on the slide was 24.3%. And our 74% to 100% result is one small support-ticket app, not a study.

The hour in five lines

1. Delete old prompt habits.
2. Fix your cache.
3. Smart model, low effort.
4. Build a real test.
5. Then let Claude climb it overnight.

The full talk is worth your hour: "Your Old Prompts Are Making Claude Worse" by Lance Martin, Anthropic, at AI Engineer. If you watch it, jump to 39:24.

Get the free cheat sheet

We put all six commands on one free page: what each one fixes and when to run it, the prompt lines to delete, the effort ladder, and the version for people who don't code. Comment FRONTIER on the video, or grab it below.

How we tested: the talk quotes and benchmark figures come from Lance Martin's 64-minute talk at AI Engineer. Our own numbers come from one run of build-eval and hillclimb on a small support-ticket app in September 2026. One small test, not a study. Independent, not affiliated with Anthropic or AI Engineer.