adhd: a parallel divergent ideation framework that prevents anchoring bias in AI agents

What it solves

ADHD addresses "premature convergence" and anchoring bias in autoregressive reasoning. In standard Chain-of-Thought or Tree-of-Thought, agents often lock onto their first idea and fail to explore a wide range of non-obvious solutions or identify potential traps in their reasoning process.

How it works

The project implements a two-phase loop with a strict mechanical separation between generation and evaluation:

1. Diverge: The system spawns N parallel, isolated LLM calls. Each call is given a specific "cognitive frame" (a vantage point prompt) and is forbidden from evaluating the result. Because these processes have zero shared context, they cannot anchor on each other's ideas.
2. Focus: A separate critic LLM call scores the ideas based on novelty, viability, and fit. It then clusters the ideas, prunes "traps" (seductive but broken ideas), and deepens the top-K survivors into detailed sketches.

Who it’s for

It is designed for developers and AI agent builders who need high-quality ideation for complex tasks such as design decisions, API surface design, fuzzy debugging, naming, and strategic planning.

Highlights

• Architectural Fix: Treats anchoring as a structural problem rather than a prompting one by using isolated parallel processes.
• Trap Detection: Specifically identifies and explains why certain common but flawed ideas should be avoided.
• Agent-Agnostic: Can be installed as a skill for various agents (Claude Code, Cursor, Windsurf, etc.) or used as a TypeScript library/CLI.
• Proven Performance: Benchmarks show significant gains in novelty and trap detection compared to single-shot baselines.