What it does ADHD is a skill (and CLI/library) for coding agents that tackles premature convergence in reasoning. It runs a two-phase loop: first, it spawns N truly isolated agent calls under different “cognitive frames” — each branch sees only its own distorted vantage point, with a hard wall between them. Second, a separate critic call scores every idea on novelty, viability, and fit, flags traps, clusters by underlying angle, and deepens the survivors into sketches with risks and first steps.

The interesting bit The isolation is architectural, not just prompting. Standard Tree-of-Thought still shares a single context window across branches, so early ideas anchor later ones. ADHD uses separate LLM calls with opposite system prompts — generator and critic are mechanically split, not politely asked to role-play in one chat. The README claims this is “an architectural fix, not a prompting one.”

Key highlights

• Ships as a skill for Claude Code, Codex, Cursor, and ~50 other agents via npx skills add
• 15 built-in cognitive frames; you can author your own
• Evaluated on 6 open-ended engineering problems against a single-shot baseline: 5 of 6 wins, with the biggest gap in trap detection (9.50 vs 1.83)
• Already adopted by multiple OSS projects: repowire, mstack, zk-flow-oss, and others
• MIT licensed; preprint published; featured in The New Stack

Caveats

• The eval methodology is self-reported: “judged by an independent LLM with a skeptical-staff-engineer prompt” — not human experts
• Cost and latency are acknowledged concerns; the “When to use” doc presumably covers this, but the README only teases it
• The 740 stars are recent and concentrated; long-term stickiness unclear

Verdict Worth installing if you regularly ask agents for design alternatives, naming help, or fuzzy debugging — the kind of work where first-idea anchoring quietly sabotages you. Skip it if you need fast, cheap, single-answer responses; the parallel calls add overhead by design.