Using Claude Fable 5 with the GOAL Framework

The GOAL Framework for Agentic Workflows

Claude Fable 5 operates as a high-level orchestrator that manages sub-agents to complete complex tasks. To maximize performance, users should adopt the GOAL framework, which shifts the interaction model from micromanaging steps to defining a desired end state.

• G (Ground in Truth): Provide the model with existing context, such as codebases, documentation, or Obsidian vaults, before asking it to reason. Use the command here is my [project] in this folder, read all of it first to ensure the agent understands the current state.
• O (Outcome, Not Orders): Define success criteria clearly rather than listing specific tasks. If the goal is vague, use a meta-prompt to force the model to interview you until it can define testable criteria.
• A (Autonomy Over the Path): Resist the urge to dictate the implementation steps. Allow the model to determine the file changes, tool usage, and build order, as it often identifies more efficient paths than the user.
• L (Loop in Proof): Require the model to verify its own work. Instruct it to open results in a browser, show before-and-after comparisons, and pause for human approval before executing irreversible changes.

Implementation and Dynamic Workflows

To execute these goals, use the /model command to select Fable and the /effort command to enable Ultra Code dynamic workflows. This configuration allows the orchestrator to spawn and manage hundreds of sub-agents that perform tasks and verify each other's output. Because Fable 5 and dynamic workflows are approximately two times more expensive than Claude 3.5 Sonnet, users should monitor token usage and set clear stop-points for human intervention.