Claude Code Lead: Stop Using AI Conservatively — Four Dimensions for Ambitious Prompting

Using AI with More Ambition

Felix Rieseberg, lead for Claude Code desktop, argues that most people use AI too conservatively. He outlines four dimensions of ambition — speed, height, depth, and breadth — and provides concrete techniques for each.

Height: Work at Higher Abstraction Levels

Instead of asking AI to execute low-level tasks, define the ultimate outcome. Rieseberg illustrates with a quarterly business review (QBR) at a fictional Acme company:

Step 1: Give the model all available data and ask: "Tell me what changed this quarter, why, and what actions we should take."

Step 2: Ask for the three most urgent decisions with pros/cons for each direction.

Step 3: Redesign the workflow: "Design a better way than quarterly slides, prove with historical data it would have caught issues earlier, and run it in parallel for a quarter."

He notes this works best when the task leads to decisions; a good signal is "if you'll need a meeting afterward, raise the goal one level." He also describes enabling Claude to keep an app crash-free by giving it access to crash reports, Git commits, and multi-OS machines for independent reproduction and verification.

Depth: Think Like Top Experts

Push the model to use the same tools, data, and methods as domain experts.

Insist on expert tools: When Claude reaches for Markdown tables, require databases and statistical models instead.

Set "unreasonable" goals: To cut load time from 10s to 5s, demand 0.2s. When the model says impossible, insist it's solvable and keep iterating.

Challenge assumed conclusions: "We think churn is an onboarding problem. Reproduce this from raw data."

Discover new knowledge: "Build a customer simulation from logs, replay the last year, find the earliest unknown churn signals, and validate on unseen data."

External validation: Never let the model grade itself. Use runnable tools (code execution, new data) for immediate feedback. Rieseberg provides a meta-prompt that asks the model to first list what top professionals do, what raw materials they use, what methods/tools they employ, and what external proof exists — then rewrite the prompt accordingly.

Breadth: Import Methods from Other Fields

Reframe nouns as verbs (from The Art of Innovation ): "Make a QBR deck" becomes "In 30 minutes, convince busy people of the causal story and decide under uncertainty." This surfaces relevant domains (journalism, analytics, motion design, audio, red-teaming). Assemble a cross-domain team of virtual specialists — analyst, journalist, editor, motion designer, audio designer, skeptic, forensic accountant — but apply their actual methods, not role-play.

All these techniques require the model to have computer access (file system, software install, data via MCP or similar).

Self-Efficacy: Upgrade Your Own Beliefs

Rieseberg connects ambition to psychological self-efficacy (Bandura). Cites Locke & Latham's 35-year goal-setting review: specific challenging goals outperform "do your best." References Patrick Collison's "Fast" list of achievements under bold deadlines. Three documented boosters:

Believe hard things are doable: George Dantzig solved two famous unsolved statistics problems because he thought they were homework.

See others break barriers: Roger Bannister broke the 4-minute mile in 1954; 46 days later others followed. Same pattern with sub-2-hour marathon (Sawe, April 2024).

Allow failure: Linus Torvalds released Linux as a hobby, "not big and professional like GNU" — yet Linux won. The Inner Game of Tennis argues over-trying degrades performance.

In an era where model capabilities jump every few weeks, updating self-efficacy is critical to attempt things we never believed possible.

Reference: felixrieseberg.com/how-to-be-more-ambitious-a-tutorial-for-humans-with-models/

Code example

最后，想学习这个项目的可以查看项目地址：