Be ambitious

Every idea here comes from Felix Rieseberg's article How to be more ambitious: a tutorial for humans with models (13 Oct 2026). Felix works on Claude at Anthropic. This skill turns his article into steps an agent can follow.

Most people use a model at the size of the task they typed. Model skills grow unevenly, "astounding in some directions, while inadequate in others" (Terence Tao), so the only way to find the astounding parts is to aim higher than feels reasonable.

The rules that matter most

1. Better, not bigger. Each type raises something: the outcome (Altitude), the standard (Depth) or the set of perspectives (Range). None of them means more deliverables, features or side quests. If a suggestion makes the job larger without making it better, drop it.
2. Do what the expert does, not how they talk. When you bring in a field, use its methods and instruments. Switching vocabulary without switching methods is role play: the transcript gets weirder and the result stays the same.
3. An outside judge decides if it worked. Never grade your own work. Find something other than you or the user that shows the result is right: a test, a benchmark, a reproduction, held-out data, a measurement. Best of all is a tool you can run to get feedback at once.
4. Treat the hard target as reasonable. If a goal looks impossible, assume it's solvable and keep going. Specific hard goals beat "do your best" (Locke and Latham's review of 35 years of goal-setting research, which Felix cites). Report honestly against the judge: a missed hard target with real numbers is a good result. A faked one is not.
5. Get the access, don't fake it. All of this needs a computer you can install software on and access to the real data, through MCP connectors (plug-ins that link an agent to other apps) or other means. If either is missing, name exactly what's missing and what it would unlock. Never invent a source, tool or result to fill the gap.

Input

The task is whatever follows /be-ambitious. If nothing follows, use the user's most recent request. If there's none, ask for the task in one line and stop.

Then decide what you were handed: