- Anthropic developer Thariq Shihipar says output quality from Claude‘s Fable 5 is limited less by the model than by users’ ability to identify their own blind spots before prompting.
- He frames prompting around four categories — known knowns, known unknowns, unknown knowns, and the critical “unknown unknowns.”
- His techniques include a “blindspot pass,” structured interviews, brainstorming design directions as HTML artifacts, and an implementation-notes.md log.
- Shihipar edited the Fable launch video entirely with Claude Code, applying the same approach in unfamiliar territory such as transcription and color grading.
What Happened
Anthropic developer Thariq Shihipar published prompting guidance for Claude’s Fable 5 model arguing that output quality now hinges on how well users recognize their own knowledge gaps before they write a prompt, as reported by The Decoder on July 4, 2026. Shihipar says Fable 5 is the first model where output quality is limited by the user’s ability to clarify their “unknowns” rather than by the model itself.
Why It Matters
The framing marks a shift in where the bottleneck sits: as models improve, the limiting factor moves from raw capability to how precisely a user can specify the problem. That reframes prompt engineering less as wording tricks and more as a structured process of surfacing what you don’t yet know.
It also comes directly from Anthropic, giving developers a first-party view of how to work with a current Claude model rather than inferring practices from trial and error. The advice targets agentic coding specifically, where a plan can look complete yet still produce the wrong result.
Technical Details
Shihipar organizes prompting around four categories: “known knowns” are what is already stated in the prompt; “known unknowns” are questions you know you have not answered; “unknown knowns” are things so obvious you would never write them down but would recognize on sight; and “unknown unknowns” are things you have not considered at all — the category he calls critical. He warns that specificity cuts both ways: too much detail risks Fable 5 rigidly following instructions even when a change of course would be better, while too little produces decisions based on industry defaults that do not fit the task. “When you don’t account for your unknowns you fail both ways,” he writes.
His pre-implementation techniques include a “blindspot pass” — asking Claude to identify your unknown unknowns, with an example prompt for adding an auth provider in an unfamiliar codebase — plus brainstorming several radically different design directions as HTML artifacts, and structured interviews in which Claude asks questions one at a time, prioritizing those whose answers would change the architecture. He treats source code as the best reference, and has Claude write an implementation plan focused on the parts most likely to change, such as data models and type interfaces, leaving mechanical refactoring for last.
Who’s Affected
The guidance is aimed at developers and agentic coders using Claude Code and Fable 5, and at the broader “vibe coding” workflow where users hand large tasks to an AI agent. During implementation, Shihipar has Claude Code keep a temporary “implementation-notes.md” file tracking decisions, and tells it to pick the conservative option on unexpected edge cases, log the deviation, and keep working. After implementation he uses “pitches and explainers” that bundle prototype, specs, and notes, and “quizzes” in which Claude generates a report and a test — he says he does not merge until he passes the quiz without errors.
What’s Next
Shihipar illustrated the approach with the Fable launch video, which he edited entirely with Claude Code despite no prior video-editing experience — learning transcription with Whisper, cutting with ffmpeg, prototyping fades with Remotion, and having Claude teach him color grading once he realized he did not know what “good” looked like. He published a visual version of the tips on a companion website. His summary of the method: “Every explainer, brainstorm, interview, prototype, and reference is a cheap way to find out what you didn’t know before it gets expensive to fix.” The advice is one practitioner’s workflow rather than a benchmarked result, but it reflects how Anthropic frames working with its latest model.
