Fable 5 prompting tips: Anthropic's guide to finding blind spots

Anthropic developer Thariq Shihipar published a set of prompting techniques for Claude's Fable 5 on Jul 4, 2026, arguing that the model's output quality is increasingly limited by users' failure to clarify their own blind spots before writing prompts. He calls out four categories of knowledge users must check for, and offers concrete practices to expose them during planning and implementation.

How should developers prompt Fable 5?

Shihipar says start by acknowledging which gaps you already see and then run a deliberate process to surface the gaps you do not. He labels four categories—Known Knowns, Known Unknowns, Unknown Knowns, and the critical Unknown Unknowns—and recommends a "blindspot pass" and other structured techniques to find what you have not considered.

Shihipar lays out several concrete techniques. Begin with a blindspot pass, where you ask Claude to identify relevant unknown unknowns for the task; he offers this example prompt: "I'm working on adding a new auth provider but I know nothing about the auth modules in this codebase. Can you do a blindspot pass to help me figure out my relevant unknown unknowns and help me prompt you better." For domains with many unknown knowns, such as visual design, he runs brainstorming and rapid prototyping with Claude, asking it to generate radically different design directions as HTML artifacts so he can react to concrete alternatives.

He also uses structured interviews, letting Claude ask prioritized questions about ambiguities that would change architecture, and treats source code as the best reference even across languages. Before coding, Shihipar has Claude create an implementation plan that focuses on parts most likely to change, like data models and type interfaces, leaving mechanical refactoring for last. He starts almost every coding session with an exploration or brainstorming phase to consciously define project scope.

During implementation he keeps a temporary implementation-notes.md file where Claude logs decisions and deviations. When edge cases arise, he instructs Claude to pick the conservative option, record the deviation, and continue. After a prototype, he bundles the work into "pitches and explainers" and runs Claude-generated quizzes: "he doesn't merge until he passes the quiz without any errors."

Shihipar demonstrates the workflow with a practical example: he edited the launch video for Fable entirely with Claude Code. Video work was new to him, so he used Claude to edit and transcribe video, consulted how Whisper handles transcription and whether ffmpeg could cut filler words and pauses, and prototyped time-controlled fades with Remotion. When color grading felt flat, he had Claude teach him the subject to expose what he didn't know about color standards instead of blindly evaluating variations.

Why does this matter?

Fable 5's gains in capability shift the bottleneck from model limitations to user framing and domain knowledge. Shihipar argues that too-specific prompts can lock the model into flawed approaches while overly vague prompts yield generic, industry-default decisions that do not fit unique tasks. He warns, "When you don't account for your unknowns you fail both ways," and shows how structured prework reduces expensive downstream rework.

The practical effect is developers can scale higher-risk tasks with Claude if they invest time up front to map unknowns, keep decision logs, and require passing quizzes before merging. That changes where developer time is spent: less on correcting model hallucinations and more on defining the boundaries and exceptions the model should respect.

What to watch

Look for whether other teams adopt blindspot passes and implementation-note practices as standard preflight checks for agentic coding. A concrete signal will be teams requiring coded quizzes or implementation-note logs before merges, or broader use of Claude Code for nontraditional tasks such as video editing where Shihipar demonstrated the approach.

Shihipar's handbook of tips and a visual version of his techniques accompany the write-up, and his example workflows—blindspot pass, structured interview, implementation-notes.md, pitches and quizzes—offer a repeatable checklist for teams using Fable 5.