ADHD Agent — Open-Sourced Skill for Claude Code
Post by Charly Wargnier — “Someone literally open-sourced an ADHD agent skill for Claude Code”
The Problem
Standard autoregressive reasoning (Chain-of-Thought) anchors hard. Once an LLM picks a direction, it stops searching.
The Solution: ADHD Cognitive Architecture
Forces divergence before convergence:
Phase 1: Diverge
- Spawns parallel, isolated calls with totally different cognitive frames
- Branches share zero context → anchoring is impossible
Phase 2: Hard Wall
- No communication between branches
- Forces exploration of radically different approaches
Phase 3: Focus (Critic Pass)
- Separate LLM pass ruthlessly scores:
- Novelty
- Concept clustering
- Flags “seductive but broken” traps before they cost engineering time
Results
| Metric | Single-shot | ADHD Agent |
|---|---|---|
| Trap detection | ~2 | ~8 |
Catches things like “streaming tokens in reverse” before wasting an hour building it.
Installation
Single npx command — works with:
- Claude Code
- Cursor
- Windsurf
- Codex
100% open source. Repo link
Related
- Built by Udit Akhouri
- Prevents premature convergence in LLM reasoning
- Anti-pattern: CoT anchoring