Anthropic engineer shares how to set more ambitious AI tasks and publishes a ready prompt

Felix Rieseberg, a lead engineer at Anthropic working on Claude, wrote a guide on how to actually set more ambitious tasks for AI and shared a ready-to-use prompt.

He notes that everyone advises users to be more ambitious with AI, but few explain how. Rieseberg highlights three approaches that go beyond simply speeding up work, and they can be applied to any AI.

Instead of asking AI to fix errors one by one, give it an end goal. Rieseberg stopped asking Claude to fix each crash separately and instead set the goal of making apps not crash at all. He provided Claude with error reports and Windows, macOS, and Linux machines so the model could test fixes on its own.

When preparing a quarterly report, first give AI all the data and ask it to identify what changed, why it happened, and what should be done. Then ask it to highlight the three most urgent decisions with their pros and cons. At an even higher level, ask AI to create something more useful than the report itself and prove with the company's historical data that it would have helped detect problems earlier.

When AI returns a small table, ask it to build a database and a statistical model instead, because that is what a leading specialist in the field would use. If a page should load in 5 seconds, set the target at 200 milliseconds. And never let AI grade its own work — use independent verification.

To bring in knowledge from other fields, assemble a team: an analyst, a journalist, an editor, and, for critical evaluation, someone betting on the company's collapse and an accounting expert. But do not just ask AI to role-play these specialists. Ask it to conduct the research they would have done.

Rieseberg also shared a ready prompt:

I need a deeper version of this task, not a bigger one.

Before rewriting my prompt:

1. Tell me what a leading specialist in this field would do that I am not currently asking for.
2. Name the most raw data or materials they would use instead of information already prepared for me.
3. Specify the concrete methods and tools they would apply.
4. Explain what evidence independent of us would confirm that the result is correct.

Then rewrite my prompt so it demands exactly that.

The prompt was turned into a skill called /be-ambitious. Just point your agent to the skill and ask it to install it: GitHub: be-ambitious.

Rieseberg's full post is also worth reading.