How to Make a Task Deeper, Not Bigger: An Ambition Skill for Claude

Asking an AI for more ambition usually makes the task bigger: more sections, more deliverables, more words. The Ambition Calibrator skill does something narrower. It decides whether a request is worth expanding at all, and if it is, it changes the inputs, methods, and proof rather than the size. It is based on Felix Rieseberg's post How to be more ambitious: A tutorial for humans with models (13 October 2026), which splits ambition into Speed, Altitude, Depth, and Range. The skill covers the last three.

When does the skill expand a request?

Only on a trigger. It expands when you ask for depth ("go deeper", "as good as you can") or hand over an outcome-type deliverable such as a report, deck, analysis, plan, or performance goal. Everything else is done as asked, with a one-line note: Ambition check: bypassed (reason).

Request	What the skill does
"run the tests"	Bypass. A short, clear command gets done without a question.
"rename tmp to retry_count in one file"	Bypass. If tmp does not exist in the file, it reports that instead of inventing an edit.
"just tidy this report"	Bypass. "Just" is a ceiling; it mentions a deeper version exists, once.
"fix the failing test"	Not expanded. Routine debugging goes to the Empirical Diagnostician skill if you have it.
"rewrite my prompt to be clearer"	Not expanded. Prompt-wording rewrites go to the Prompt Complexity Triage skill if you have it.
"make the quarterly review deck from these exports"	Expands, then stops at a confirmation gate.
"go deeper on making this endpoint faster"	Expands.

What does "deeper, not bigger" mean in practice?

An expansion may change four things: the inputs, the methods, the target, and the proof required. It may not add deliverables you did not ask for. A new index, migration, cache, scheduled job, or dependency counts as outside the task, so it goes into a follow-up list that needs your approval.

How does the Altitude step work?

The skill writes a three-rung ladder: the base task, the direct result, and the outcome the task is a step toward. For a review deck, that is the deck, then a decision-ready review, then problems caught early. It skips the ladder for creative work whose value is the artifact itself, and it lists missing prerequisites (a data feed, a way to reproduce and verify) instead of claiming an outcome will be maintained after one run. It will not set up anything recurring without your consent.

What does the Depth step ask?

Four questions, answered in writing before any rewrite:

1. What would a top professional do here that the request does not ask for?
2. What is the rawest material they would start from, instead of what has already been digested? The skill has to inspect what exists before naming it.
3. Which specific methods and instruments apply?
4. What, other than the assistant or you, would show the result is right?

Your stated target stays the acceptance bar. The skill can add a harder number labeled stretch, and if you gave no target it proposes one labeled proposed. The article's own one-paragraph prompt for this step is quoted inside the skill, with credit.

What is Range, and why not use personas?

Range borrows methods from adjacent fields. The skill first rewrites the noun as a verb (a quarterly deck becomes getting busy people to believe a story about cause and effect and to decide under uncertainty), then picks up to five lenses, such as an analyst who reconciles totals or a short seller who looks for the weakest claim. A lens only counts if it runs a concrete check. Asking the model to write like a short seller changes the vocabulary; asking it to do what a short seller does changes the work. Lenses run inline. Spawning subagents is opt-in.

What is the outside judge?

The skill never grades its own work. It picks, in order of preference: an executable check (tests, a linter, a benchmark, a query on the raw data), a held-out check, a reviewer with fresh context, or you with specific questions to answer. It shows the command and an unedited excerpt of the pass and fail lines. When no outside judge exists, which is common for decks, plans, and creative work, the verdict is UNVERIFIED: no external judge available, followed by what would verify it. If the judge fails three times, it stops and hands you the output. It does not change the judge to make it pass.

When does it ask before running?

When the rewrite widens scope, costs materially more, needs new access or installs, writes outside the working files, reads large or sensitive data, or is hard to undo. If it is unsure whether a rewrite crosses one of those lines, it asks. In a non-interactive run it returns the rewrite marked awaiting confirmation and does not execute.

What has been tested?

Two read-only dry runs with fresh agents. The first, on two requests, showed the original version expanding too eagerly, so the trigger rules were rewritten. The second used six prompts with pass conditions fixed in advance: "run the tests" bypassed without a question, "fix the failing test" produced no rewrite block, a rename of a variable that did not exist was reported, the deck request expanded and stopped at the gate, "go deeper" on an endpoint expanded, and a prompt-wording request deferred. All six matched; the last wording edit was not re-run. This is a smoke test on a handful of prompts. It does not show that deeper rewrites beat plain prompts, and the skill makes no such claim.

How do I install it?

1. Download ambition-calibrator.md.
2. Save it as .claude/skills/ambition-calibrator/SKILL.md in your project, or in ~/.claude/skills/ambition-calibrator/SKILL.md to use it everywhere.
3. Ask Claude for a report, an analysis, or "go deeper" on something. It loads from the skill's description.

The skill is a set of instructions Claude follows, with no backend call and no account. More skills are in the GitHub repository.

FAQ

Is this the same as asking for a longer answer?

No. It forbids adding deliverables. It changes what the work is built from and how it is checked.

Does it work with models other than Claude?

The file is plain instructions, so any agent that loads skill-style files can follow it. The dry runs used Claude.

Why does it so often bypass?

Expansion costs time, money, and your attention. A rename or a test run does not benefit from it.

What if there is no way to run a check?

It says so. The result is labeled UNVERIFIED and you get the exact questions a human reviewer should answer.

Related: Prompt Complexity Triage decides how much to change a prompt's wording, and Subagent Dispatch Economics decides when a subagent is worth its cost. The ambition framework comes from Felix Rieseberg.