Most single-shot or shared-context Chain/Tree-of-Thought methods suffer from premature convergence: early lines of reasoning anchor later outputs. ADHD treats anchoring as an architectural problem rather than a prompting trick — it deliberately spawns N isolated, frame-shifted reasoning processes, forbids cross-branch leakage during divergence, and runs a separate critic pass to score, cluster, prune traps, and deepen survivors.
What Sets It Apart
- Mechanical generator→critic split: the diverge phase runs isolated LLM calls under different cognitive frames (no shared context), and the focus phase uses a separate critic to evaluate novelty, viability and traps. This separation reduces anchoring in a way prompting alone cannot.
- Practical, integrable skill: distributed as a runnable agent skill with CLI and TypeScript library hooks, plus a preprint and reproducible evals. Early adopters and integrations show the pattern shipping in multiple OSS projects rather than being a one-off paper artifact.
- Measured gains on open-ended engineering problems: independent LLM judges reported big improvements in breadth, novelty and trap detection (example headline numbers: breadth ~9.0, novelty ~7.8, trap detection ~9.5 vs much lower single-shot baselines), at a predictable cost in time and compute.
How It Works (short)
- Diverge: choose N cognitive frames and spawn N isolated agent calls; branches never see each other and are prevented from evaluating during generation.
- Focus: a separate critic call scores each idea (novelty/viability/fit), clusters similar angles, flags traps with reasons, prunes low-quality branches, and deepens top survivors into actionable sketches.
Who It's For and Tradeoffs
Great fit if you need broad, non-obvious ideation or robust trap-detection for design, API surface decisions, naming, strategy, or any "give me a few ways to…" prompt. It integrates easily into agent ecosystems (Claude Code, Codex, CLI) and exposes frames/APIs for customization. Look elsewhere if latency, deterministic minimal-cost outputs, or single-answer consistency are primary constraints: ADHD typically multiplies runtime and token costs (authors report ~2× compute/time) to gain breadth and safety, so it’s a deliberate cost/benefit tradeoff for creativity and risk-aware planning.
