Using Tree of Thought Agents to Improve AI Ideation

The Breakthrough

The ADHD skill replaces standard linear AI ideation with a Tree of Thought architecture that forces agents to explore multiple, isolated problem-solving paths, preventing the common issue where models produce redundant, safe variations of the same idea.

How It Works

• Isolated Branching: The skill spins up multiple sub-agents that operate in separate context windows, ensuring that ideas from one branch do not influence or bleed into others.
• Frame-Based Ideation: Each agent is assigned a specific 'frame' from a library, which acts as a unique lens or perspective to guide its reasoning process.
• Skeptical Critique: A dedicated critic agent evaluates all generated ideas based on three metrics: novelty, viability, and fit. It uses a 'senior engineer' persona to filter out obvious or weak suggestions.
• Pre-Check Logic: To conserve tokens, the skill runs a pre-check to determine if a problem is truly open-ended. It skips execution if the request is for a 'quick' or 'standard' answer, or if there is only one objectively correct solution.
• Non-Obvious Prioritization: The system explicitly ignores the first three potential answers, as these represent the most common patterns in the training data, and instead focuses on non-obvious, high-value outputs.

Practical Application

• Test-Driven Development (TDD): By linking a PRD and technical specification in claude.md, users can generate comprehensive test strategies that cover edge cases and performance paths often ignored by standard agents.
• Pre-Launch Audits: The skill can evaluate live products or features to identify churn risks and UX gaps, effectively surfacing discrepancies between original PRD requirements and the actual implementation.

Context

Standard AI agents struggle with creativity because they rely on the most probable patterns in their training data, leading to repetitive, 'safe' outputs. When asked to brainstorm, they often fill their context window with noise, causing the quality of reasoning to degrade. This skill addresses that by forcing divergence through structural isolation, providing users with a ranked report of strategies—scored in shorthand like N9 (Novelty 9), V8 (Viability 8), and F10 (Fit 10)—which the user can then instruct the agent to implement.