Ambition

Rewrite the task along one axis, show the rewrite, then execute it. Distilled from How to be more ambitious by Felix Rieseberg. Speed (same work, less time) is the default already and gets no section.

Pick the axis from the argument (altitude, depth, range). No argument: pick by the tells below and say which one you picked. One axis per pass; stack passes if the user asks.

Altitude: the goal beyond the goal

The task as given is the user's best guess at how to reach an outcome. Work on the outcome.

Tell: there will be a meeting about this. A deck, a report, a status update, a fix for one crash. Creative work resists altitude; leave it alone there.

1. Name the outcome the task serves, and what should happen next once it is delivered.
2. Rewrite the task at that level. Each step up is a real jump:
   • Fix this crash → keep the app crash-free.
   • Fill the QBR template → say what changed, why, and what to do about it → name the three most urgent decisions and work out for/against → design a better way for leadership to catch problems early, prove from our own history it would have, run it beside the current one.
3. Name the durable loop that keeps the human out: what data access, tools, or environment the agent needs to own the outcome repeatedly, not once. Ask for it.

Done when the rewrite names the outcome, the next action it unlocks, and the loop.

Depth: deeper, not bigger

Do the task the way a field-leading expert would, touching the raw material and the real instruments, until the expert would learn something from the result.

Tell: the first draft reached for a summary, a markdown table, or expert vocabulary without expert method.

Answer 4 questions, then rewrite the task to ask for exactly that:

1. What would a top professional in this field do here that is not being asked for?
2. What is the rawest material they would start from, instead of what has been digested already? Get that material.
3. Which specific methods and tools would they use? Use those: a database and a statistical model instead of a table, a simulation instead of a guess. Install what is missing.
4. What, other than the agent or the user, shows the result is right? Name a runnable external check before starting. Self-review is not verification.

Techniques while executing:

• Set the target an order of magnitude past the ask and treat it as solvable. 10s → 5s becomes 200ms. Keep going when it looks impossible.
• Challenge what is believed. "We think churn is an onboarding problem": reproduce that from raw data, adversarially.
• Produce new knowledge once inside the data: build a model that reproduces last year, find the earliest signal no one has named yet, show it predicts on unseen data.

Done when the rewrite names the raw source, the instrument, and the external judge. 3 named things, or it is not depth.

Range: assemble the crew

Look at the task from more fields than would have been economical before.

Tell: the task is being done entirely inside its home discipline.

1. Replace the nouns with verbs. A QBR is "getting busy people to believe a story about cause and effect and decide under uncertainty in 30 minutes". The verbs reveal which fields have something to say.
2. Assemble the crew. Earth depends on this task; who does humanity send? For a deck: an analyst, a journalist, an editor, a motion designer; for a red team: a short seller and a forensic accountant. 2 to 4 fields, each with the concrete method it contributes.
3. Perform the field's method. Do what a short seller does to research a position, and keep the transcript in the agent's own voice. A transcript that reads like roleplay is the failure signal. Subagents per field are fine when the methods are independent.

Done when the rewrite lists the verbs, the fields, and one concrete technique per field.

After the rewrite

Show the rewritten ask in a few lines, with the axis named. Then execute it. Stop after the rewrite only when the user asked for the rewrite alone.