Prompting was the first skill anyone had to learn to work with AI, and it is still the one that separates good output from useless output. Margot van Laar , an applied AI engineer at Anthropic, walked a room at the Code with Claude conference in London through fixing a broken customer support prompt live, one failure at a time.

Her first move is to define what better looks like. Teams need evaluations, she said, to know whether “a change to our prompt is actually correlating to an improvement” rather than simply producing a different answer.

This guidance applies across ChatGPT and Claude. Test your prompt against a few cases you care about, then change one thing and check if it helped. Guessing is not a method. Here are the five fixes van Laar reached for most.

How to get better output from ChatGPT and Claude: lessons from the Code with Claude conference

Separate the parts of your prompt

Prompts grow into one messy block over time. Role, rules, tone, and data all blur together, and the model loses the thread. "If you're reading a prompt and you can't tell guidelines from policy from data, most likely the model isn't able to either," van Laar said. The first fix is structure.

Break the prompt into labelled sections. Put the role in one place, the rules in another, the data in another, and the task last. Van Laar cleaned up the structure of her example before changing a single instruction, and the output improved on its own. Clear sections do work that clever wording cannot.

Delete instructions that no longer help

Old instructions turn into liabilities. Van Laar's support bot refused to give a customer information they were entitled to, because someone had once written a rule telling it to point customers to a web page instead of risking a wrong answer. "Instructions like these have now become redundant and are actually being overfitted to," she said. The model had been built to follow them, so it followed them too well.

Go through your saved prompts and cut the patches you wrote for weaker models. Newer models follow instructions more literally, so a defensive line that helped last year can break things now. "The model can withhold information that it actually has access to," van Laar said. Keep only the instructions that still apply.

Give it a tool instead of telling it to try harder

A stern instruction does not add a skill. Van Laar's bot kept botching a billing calculation, and the prompt's answer was to insist, in capital letters, that it calculate correctly. "Instructions don't add capability. Telling the model it's critical to do a calculation right doesn't make it better at mental math," she said. The fix was to hand it a calculator.

Watch for the step your model gets wrong again and again. Maths, lookups, anything exact. Telling it to be careful will not move the result. Connect it to a tool that does the job, or run that step outside the model and feed back the answer. Robots don't respond to pep talks.

Show both sides of every tradeoff

One-sided instructions push the model too far one way. Van Laar's bot would not escalate a billing error to a human, because the prompt told it escalating costs eight dollars and counts against the team, with nothing about the upside. "We need to remember to state both sides of the tradeoffs, because our models are becoming better themselves at making those tradeoffs," she said.

When you ask a model to weigh something, give it the full picture. Name the cost and the benefit, the risk of acting and the risk of doing nothing. A model that only hears one side will optimise for that side every time. Trust it with the whole decision.

Split one big prompt into three

Hard tasks break a single prompt. When van Laar built an agent to schedule shifts, one large prompt failed every test. So she split the job into three independent prompts that generate a draft, check it for errors, and repair it. "We have three very simple prompts, but they're now running independently rather than trying to do everything," she said.

Do the same with any complex task. Write one prompt that produces a first draft, a second that checks the draft against your rules, and a third that fixes what the second one flagged. The split solved every test case van Laar had been failing, using fewer tokens and less time.

Make ChatGPT and Claude give you better output

What turns a mediocre prompt into a reliable one? Structure, fewer instructions, the right tools, both sides of a tradeoff, and the willingness to break one prompt into three. The method underneath all five is the same. Test your prompt against a handful of cases you care about, then fix one failure at a time. Start with the prompt you use most and clean it up, and remember this guidance for every prompt you use this week.

Get more insights on using AI for your business in my AI playbook .