Finding Your Unknowns: Practical Patterns for Collaborating with Claude Fable 5
The author shares iterative patterns for discovering unknowns when collaborating with Claude Fable 5, including blindspot scans, prototyping, interviews, reference materials, implementation plans, implementation notes, proposals, and quizzes, illustrated with a case study of shipping a launch video.
Using Claude Fable 5 teaches an old lesson: the map is not the territory. The map represents the work to be done — prompts, skills, and context given to Claude. The territory is where work actually happens: the codebase, the real world, and its constraints. The gap between map and territory is what the author calls unknowns . When Claude hits an unknown, it decides based on its best guess of the user's needs. More completed work means more potential unknowns.
Fable is the first model where the author felt work quality was limited by the ability to clarify unknowns. Upfront planning is not always enough; unknowns can surface deep in implementation or reveal that a completely different approach is needed. Collaborating with Fable is an iterative process of discovering unknowns before, during, and after implementation.
Know Your Unknowns
When bringing a problem to Claude, the author breaks unknowns into four categories:
Known knowns: What the prompt explicitly states — what the user tells the agent they want.
Known unknowns: What the user hasn't figured out but knows they haven't figured out.
Unknown knowns: Things so obvious the user would never write them down but would recognize instantly upon seeing them.
Unknown unknowns: What the user hasn't considered at all — knowledge they don't know they lack, or how good a solution could be.
The best agent programmers have relatively few unknowns. Observing experts like Boris or Jarred write prompts shows meticulous clarity about what they want, with high alignment to both codebase and model behavior. Yet they also assume unknowns exist. Reducing and planning for unknowns is the core skill of agent programming, and it can be improved through collaboration with Claude.
Help Claude Help You
Guiding Claude is a subtle balance. Too specific, and Claude follows instructions even when a pivot would be better. Too vague, and Claude defaults to industry best practices that may not fit the task. Without considering unknowns, both failure modes appear. Claude can help discover unknowns faster: it searches codebases and the internet rapidly, knows more on most topics, and iterates from failure quickly. The most important step is giving Claude context about the starting point — where the user's thinking stands, their experience with the problem and codebase — so it can collaborate as a thinking partner. The author notes HTML artifacts are the best way to visualize and express ideas in almost all these scenarios.
Before Implementation
Blindspot Scan
When starting work in a new codebase area or unfamiliar domain (e.g., iterating a design), the user likely has many unknown unknowns . They may not know what questions to ask, what good looks like, what historical work exists, or what pitfalls to avoid. Ask Claude to perform a "blindspot pass" to surface unknown unknowns and explain them. Providing context about who the user is and what they know is crucial.
Example prompts:
"I'm adding a new auth provider but know nothing about this codebase's auth module. Can you do a blindspot pass to find relevant unknown unknowns and help me write a better prompt?"
"I don't know what color grading is but need to color grade this video. Can you teach me the unknown unknowns of color grading so I can prompt better?"
Brainstorming and Prototyping
When working in domains with many unknown knowns — standards the user can only define after seeing them — the author has Claude brainstorm and prototype. Identifying unknown knowns early in prototyping is valuable because discovering them during implementation can be costly; a small spec change may require a vastly different code implementation, and rolling back agent changes can be harder.
For example, the user might want to see a button in a framework without wiring backend routes or maintaining extra frontend state. Visual design is hard to articulate but recognizable on sight. In such cases, the author asks for several design directions for an artifact. Almost every coding session begins with an exploration or brainstorming phase to scope the project with clear intent. Claude often finds high-value directions the user would miss, though sometimes it misses the forest for the trees. Brainstorming prevents scoping too narrowly or too broadly.
Example prompts:
"I want a dashboard for this data but have no visual taste and don't know what's possible. Make an HTML page with 4 very different design directions so I can react."
"Before wiring anything, make a standalone HTML file mocking the new editor toolbar with fake data. I want to react to layout before you touch the real app."
"Here's my rough problem: users drop off after onboarding. Search the codebase and brainstorm 10 intervention points, from lowest cost to most ambitious. I'll tell you which resonate."
Interview
After sufficient brainstorming, unknowns often remain. The author has Claude interview them around any unknowns or fuzzy areas. When asking Claude to interview, provide context about the problem to guide its questions.
Example prompt:
"Ask me one question at a time around any ambiguities, prioritizing questions whose answers would change the architecture."
Reference Materials
Sometimes the user cannot describe what they want in detail — lacking vocabulary or facing excessive complexity. The best answer is reference materials. While diagrams, docs, or images help, the absolute best reference is source code . If a library implements something the desired way, or a design component is admired, point Fable at the folder and tell it what to examine, even if it's in another language.
Claude Design works similarly: point it at a website module and it reads the underlying code, not just screenshots, providing richer detail on markup, structure, and actual construction.
Example prompt:
"This Rust crate in vendor/rate-limiter implements the exact backoff behavior I want. Read it and reimplement the same semantics in our TypeScript API client."
Implementation Plan
When ready to implement, the author asks Claude to produce an implementation plan for review, focusing on parts most likely to change: data models, type interfaces, or UX flows. This surfaces decisions the user may need to adjust.
Example prompt:
"Write an implementation plan in HTML, but lead with decisions I'm most likely to tweak: data model changes, new type interfaces, and anything user-facing. Put mechanical refactors last; I trust you on those."
During Implementation
Implementation Notes
Once satisfied with the plan, the author starts a new session with relevant artifacts (spec, prototype) and asks the agent to implement. Regardless of planning, unknown unknowns lurk. The agent may discover an edge case in the code requiring a different approach.
The author requires Claude Code to maintain a temporary "implementation-notes.md" (or .html) file recording decisions made, so learning carries into the next attempt.
Example prompt:
"Maintain an implementation-notes.md file. If you hit an edge case forcing a deviation from the plan, pick the conservative option, log it under 'Deviations', and keep going."
After Implementation
Proposals and Explanation Docs
A critical part of delivery is gaining buy-in and approval. Building proposal and explanation artifacts into final documentation helps:
Accelerate reviewer understanding when they share the same initial unknowns.
Speed approval when experts want confirmation that expected unknowns and common failure modes were considered.
Example prompt:
"Package the prototype, spec, and implementation notes into a doc I can drop in Slack for buy-in. Lead with a demo GIF."
Quiz
After a long session, Claude may have done more than the user realizes. Reading code diffs gives only shallow understanding because behavior often depends on existing code paths. Having Claude quiz the user after providing context helps internalize what happened. The user only merges after passing the quiz perfectly.
Example prompt:
"I want to ensure I understand everything in this change. Give me an HTML report explaining the changes with context, intuition, what was done, etc., and put a quiz at the bottom that I must pass."
How These Methods Connect: Shipping Fable
The Fable launch video was edited entirely by Claude Code. For the author, this was a new domain with no expertise. The process began with what was known: Claude can edit and transcribe video via code, but accuracy was uncertain. The author asked Claude to explain how Whisper-style transcription works and whether ffmpeg could accurately cut filler words like "um" or long pauses.
The author wanted a UI synced word-for-word with speech but wasn't sure it was feasible, so Claude built a prototype video using Remotion and the transcript to test viability.
The final video looked washed out — the author knew this was a color grading issue but didn't truly understand color grading. The first attempt asked Claude for several versions to choose from, but the author realized they didn't know what "good" looked like for color grading. They pivoted to having Claude teach them color grading to uncover unknowns.
A deeper explanation of this process is available via a linked resource.
Match the Map to the Territory
Stronger models enable more with the right methods. When a long-horizon task returns a wrong result, it often signals the need to spend more time defining unknowns or creating an implementation plan that lets Claude improvise within unknowns.
Every explanation doc, brainstorm, interview, prototype, and reference material is a low-cost way to discover what you didn't know before fixing it becomes expensive. So on the next project, start by having Claude help you find your unknowns.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
