Claude Engineer Reveals: Stop Using AI Timidly — Master Ambition with Height, Depth, Breadth

Anthropic's Claude engineering lead Felix Rieseberg argues that users are too timid with AI — they merely ask AI to "act as an expert" without changing the underlying method. He breaks "ambition" into four dimensions: speed (doing same work faster), height (raising abstraction to the true goal), depth (using top practitioners' tools and raw data), and breadth (borrowing methods from other fields). Speed is familiar; he focuses on the other three.

Height: Raise the Task to Its True Purpose

Instead of "fix this crash," aim for "keep the app from crashing." Before acting, ask: what should happen after this task? Who decides, and what decision will they make? Felix illustrates with a quarterly business review (QBR):

First, have AI analyze what changed, why, and what to do — not just fill a template.

Then, have AI identify the three most urgent decisions, with pros/cons for each.

Finally, challenge the QBR format itself: design a better way to surface problems earlier, run it in parallel for a quarter.

Rule: raise height only for tasks that require discussion and follow‑up action; pure creative tasks don't need it. Prerequisite: give AI access to data and tools (crash reports, git commits, test machines) so it can self‑verify.

Depth: Work Like a Top Practitioner, Not an Actor

"Act as an expert" only changes vocabulary. Depth means forcing the model to use the actual tools and raw materials a top professional would. Felix gives five tactics:

Insist on industry‑standard tools — e.g., databases and statistical models, not Markdown tables.

Set an "unreasonable" target — demand 200 ms load time instead of 5 s; the model must rethink architecture, not just compress images.

Challenge assumed facts — ask AI to reproduce your hypothesis from raw data ("verify if onboarding really causes churn"), not confirm it.

Produce new knowledge — build a simulation, find an unknown early‑churn signal, and validate it on unseen data (train/test split).

Use external verifiers — unit tests, benchmarks, lint, type checks. Never let the model judge itself.

Felix's go‑to depth prompt (the article's most valuable artifact):

I want a deeper version of this task, not a bigger one. Before rewriting it:
(1) say what a top professional in this field would do here that I'm not asking for;
(2) name the rawest material they would start from instead of what has already been digested for me;
(3) name the specific methods and tools they would use;
(4) say what, other than you or me, would show the result is right.
Then rewrite my prompt to ask for exactly that.

Breadth: Import Methods from Other Domains

Convert nouns to verbs: a QBR isn't "a slide deck" but "make busy people believe a causal story and decide under uncertainty." That invites journalists (storytelling) and investors (decision under uncertainty). Assemble a dream team for the task (analyst, journalist, editor, motion designer, audio designer) plus a "red team" (short‑seller, forensic accountant) — but only borrow their methods , not their personas.

Implementing the Framework as a Reusable Skill

The author packaged Felix's framework into a Claude Code Skill ( /ambitious) with four steps:

Height — ask user what should happen after the task; raise if worthwhile.

Depth — run the four depth questions, set an ambition target one order of magnitude higher.

Breadth — noun‑to‑verb, assemble 3‑5 roles (including a critic), list each role's method only.

Output plan — present rewritten task, tools, ambition target, external verification, and required user inputs; wait for confirmation.

Execution rules: verify after each phase with the defined external check; try at least two approaches before reporting failure; do not expand scope.

Real‑World Comparison (Opus 5.5, same prompt ± Skill)

3D Solar System : Vanilla — ~7.5 min, 85k tokens — planets orbit decoratively. With Skill — ~53 min, 330k tokens — real NASA ephemeris (max 0.20° error), click‑to‑fly camera, live data panel, accurate eclipse simulation.

3D Programmer's Study : Vanilla — ~7.6 min, 83k tokens — cartoon style, neon signs, rubber duck. With Skill — ~28 min, 210k tokens — realistic lighting (lamp, monitor, moonlight), 110 real book titles, keyboard keys press in sync with on‑screen typing.

Personal Homepage : Vanilla — ~4 min, 57k tokens — hand‑drawn title, animated flowchart. With Skill — ~6 min, 93k tokens — verified article counts, zero external deps, mobile‑safe.

Skill version costs 4‑7× time/tokens but guarantees the AI defines "good" and proves it. Use only for tasks worth doing well.

Prompt Templates to Replace Timid Requests

Instead of "fix this crash":

Keep our app crash‑free. Here are crash reports, git commits, test machines — you reproduce, fix, and verify.

Instead of "optimize to 5 s":

Get first‑screen load under 200 ms. This is achievable; if a path fails, try another.

Instead of "analyze churn as a senior analyst":

We think onboarding causes churn. From raw data, reproduce or refute that. Use databases and statistical models, not just tables.

Instead of "make a QBR deck":

QBRs exist to surface problems early. Design a better method, prove it with our history, run it alongside the old process for a quarter.

Instead of "let me know when done":

After each step, run tests / screenshot diffs — prove it works, don't say "looks fine."

If unsure how to frame it, prefix your request with /ambitious and let the Skill do the thinking.

Reference: Felix Rieseberg's original X post (https://x.com/felixrieseberg/status/2110023283169247623) and blog article "How to be more ambitious: A tutorial for humans with models" (https://felixrieseberg.com/how-to-be-more-ambitious-a-tutorial-for-humans-with-models/).