How to Build Projects With Claude Fable Goals

TL;DR

Structure each Claude Fable project around a task, reason, outcome, constraints, and measurable verification condition, then use the /goal command to keep the work running until that condition is checked. The demonstrated projects span reinforcement learning, browser games, YouTube research, graded social content, and recurring A/B tests, with success defined through concrete outputs or performance thresholds.

Transcript

In this video, we're going to talk about six insane projects to learn CloudFable and the /goal command. We're going to build 3D simulations, interactive dashboard, 247 AB tests, plus [music] three other projects that you can do with CloudFable, and the /goal command. Starting with building a 3D simulation using Genesis World, for example, here, thi... Read More

Key Insights

• A strong project prompt is composed of five parts: the task, the reason for doing it, the desired outcome, the constraints, and the verification step. This structure gives Claude both strategic context and an observable definition of completion.
• The /goal command sets a condition that Claude checks before stopping. A goal can contain the complete project prompt, while /goal clear removes an incorrect or unwanted goal so the user can replace or stop it.
• The spider simulation uses reinforcement learning in Genesis World to train a robot to climb or jump upstairs. Its deliverables include the codebase, trained model, and video, while successful learning requires reaching the top stair in at least eight of ten evaluations.
• The simulation workflow includes checking the local environment, installing required tools, finding or creating a spider asset, smoke-testing Genesis, building the stair environment, training the spider, and evaluating its performance against the stated success rate.
• The dragon project produces an in-browser fantasy game with simple keyboard controls, boosting, turning, diving, and fire breathing. Its goal requires the player to start, fly, complete an objective, and play without encountering a major bug.
• The YouTube visualization connects to an Apify account and examines titles from leading creators. It displays topic frequency, median views, common numbers and phrases, linked example videos, breakout topics, and searchable title data organized by concept, format, views, and channel.
• The content workflow installs a plugin containing post-writing and post-grading skills. It creates seven brand-aligned posts within platform character limits, evaluates every post on a zero-to-ten scale, and repeatedly improves each draft until its score exceeds eight.
• The continuous A/B testing workflow sends traffic evenly between two ManyChat messages, tracks clicks with tags, keeps the winning version, and rewrites the losing hook every four hours. It tests angles such as curiosity, exclusivity, urgency, and social proof while distinguishing meaningful differences from noise.

Questions & Answers

Q: How do you structure a Claude Fable project prompt?

A Claude Fable project prompt should identify five elements: the task, the reason the work matters, the expected outcome, the constraints, and the verification method. The reason supplies higher-level context, while the outcome names concrete deliverables. Constraints limit tools or scope, and verification provides a measurable condition that Claude can check before declaring the project complete.

Q: How does the /goal command work in Claude Code?

The /goal command sets a condition that Claude must check before stopping its work. The user can enter /goal and then provide the full project prompt or goal condition. Claude can proceed through installation, construction, testing, and evaluation steps until it verifies completion. If the goal is wrong or no longer wanted, the user can enter /goal clear.

Q: How can Claude train a spider robot to climb stairs?

The prompt asks Claude to use Genesis World and reinforcement learning to create a spider robot that can climb or jump upstairs. Claude checks the environment, installs the needed tools, creates or finds the robot asset, smoke-tests Genesis, builds the stair environment, and trains the policy. The requested deliverables are code, a trained model, and a verification video.

Q: How is success verified for the spider simulation?

Success is defined as the spider reaching the top stair in at least eight of ten evaluation attempts. This threshold is included directly in the goal so Claude has an observable stopping condition. A video also supports verification by showing whether the trained spider actually performs the intended climbing behavior instead of merely producing code or a model file.

Q: How can Claude Fable create a browser-based dragon game?

The game prompt requests a dragon-flying experience in a fantasy environment with simple controls and the feeling of riding on top of the dragon. The resulting browser game includes climbing, diving, left and right turns, boosting, and fire breathing. Players fly through rings and destroy beacons, while verification checks that the game starts, flies, completes a goal, and runs without major bugs.

Q: What does the YouTube title visualization analyze?

The interactive visualization analyzes top-performing videos from leading YouTubers to identify the psychology of interest in their titles. It charts how frequently topics appear and the median views associated with them. Users can inspect dollar figures, numbers, phrases, breakout topics, and linked examples, or search all titles to compare concepts, view counts, formats, and channels.

Q: How can Claude generate and grade a week of social content?

Claude can install a plugin containing a post-writer skill and a post-grader skill, then create seven posts for the following week. The prompt requires every post to respect the relevant platform character limit and match the user's brand voice. Each draft is scored from zero to ten and revised until every post receives a score above eight.

Q: How can Claude run continuous A/B tests on marketing copy?

Claude can use browser access to divide traffic evenly between two message versions, measure clicks, compare click-through rates, retain the winner, and replace the losing version with a new hook. In the ManyChat example, tests repeat every four hours, clicks are tracked with tags, and each round tests angles such as curiosity, exclusivity, urgency, and social proof while checking whether results reflect a real difference or noise.

Summary & Key Takeaways

• The core workflow combines Claude Code, Claude Fable, and the /goal command. Each project prompt specifies the task, its motivation, the required outcome, operating constraints, and a verification test. The goal condition gives Claude a measurable requirement to check before stopping instead of treating initial code generation as completion.

• The projects include training a spider robot in Genesis World, creating a browser-based dragon game, and building an interactive analysis of top YouTube titles. Their verification methods differ: the spider must succeed eight times out of ten, while the game must support starting, flying, completing a goal, and playing without major bugs.

• The marketing examples apply the same workflow to content production and continuous optimization. Seven posts are revised until a post-grader score exceeds eight out of ten. A separate ManyChat experiment splits traffic evenly, compares click-through rates, retains the winner, rewrites the loser, and repeats every four hours.