Anthropic’s new field guide for Claude helps users improve AI outputs by identifying and addressing various types of unknowns throughout the project lifecycle—before, during, and after building—using techniques like blind spot analysis, prototyping, and iterative documentation. This structured approach empowers users to provide clearer prompts, manage complexities effectively, and leverage AI to enhance understanding, stakeholder communication, and continuous skill development.
The video introduces Anthropic’s new field guide for Claude, designed to help users break free from the frustrating cycle of prompting, re-prompting, and still not getting the desired AI output. The core concept revolves around identifying and addressing the “unknowns” in any AI-driven project. The guide emphasizes the difference between the “map” (your initial prompt and plan) and the “territory” (the real-world complexities and unknown challenges that arise during execution). By uncovering these unknowns early, users can provide better context to Claude, leading to more accurate and useful outputs.
The guide categorizes unknowns into four types: unknown knowns (what you know and include in your prompt), known unknowns (gaps you are aware of), unknown knowns (things you know but don’t realize enough to plan for), and unknown unknowns (things you don’t know you don’t know). Experience traditionally helps uncover these unknown unknowns, but AI now makes it easier to surface and address them before starting a project. The goal is to evolve from vague, unclear prompts to well-informed, precise instructions that reflect a deeper understanding of the task and its potential pitfalls.
Anthropic’s guide breaks the AI project workflow into three phases: before you build, while you build, and after you build. In the “before you build” phase, users perform a “blind spot pass” with Claude to identify gaps in their knowledge, brainstorm and prototype ideas quickly to visualize concepts, and let AI interview them to clarify design and build choices. Providing examples to Claude is also crucial, as it helps the AI understand the desired outcome more effectively than verbal descriptions alone.
During the “build” phase, the guide recommends starting fresh with a clean context window and a well-prepared plan, incorporating prototypes and insights gained earlier. Users should expect and log edge cases or surprises in a markdown file to learn from them and improve future builds. This iterative documentation helps build expertise and reduces unknowns over time, making subsequent projects smoother and more predictable.
Finally, in the “after build” phase, the focus shifts to gaining stakeholder buy-in by quickly generating supporting artifacts like guides, slide decks, or prototypes using Claude. Additionally, users should have Claude quiz them on the project to solidify their understanding and become true domain experts. This approach ensures that AI augments human intelligence rather than replacing it, empowering users to deliver better outcomes and continuously improve their skills through a structured, AI-assisted workflow.