ADHD: A Tree-of-Thought Skill That Attacks Premature Convergence in Coding Agents
ADHD — a skill for coding agents. Tree-of-thought with pruning, built on the Claude & Codex Agent SDK. Fans out parallel divergent thoughts under different cognitive frames, scores, prunes traps, deepens the survivors. The no-brainer skill for creative and interdisciplinary work.
At a glance
- What is it?
- ADHD is a TypeScript skill for Claude Code and other agents that fans out isolated reasoning paths under distorted cognitive frames to avoid the anchoring that plagues linear chain-of-thought. It is aimed at creative and interdisciplinary tasks, not routine coding.
- Who is it for?
- ADHD is for engineers who regularly ask agents open-ended questions like design, naming, strategy, or fuzzy debugging and who are willing to pay extra tokens for breadth and trap detection. It is not for simple, well-specified coding tasks where a single pass is cheaper and faster.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 3 days ago.
- What is it written in?
- Mainly TypeScript, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The Problem: Premature Convergence in Autoregressive Reasoning
The README frames ADHD as an architectural fix for a specific failure: linear chain-of-thought anchors on whatever the model says first. Tree-of-thought widens the search but still walks a single shared context, so the anchoring persists across branches. ADHD treats this as a structural issue, not a prompting issue. It spawns N isolated reasoning processes under deliberately distorted cognitive frames, with zero shared context during divergence, then runs a separate critic pass to score, cluster, prune traps, and deepen the survivors. The intended use cases are design decisions, fuzzy debugging, naming, API surface design, strategy, and any prompt shaped like 'give me a few ways to...'.
How It Works: Isolated Frames and a Critic Pass
The mechanism is visible in the README's side-by-side eval. For a CLI hang problem, ADHD spawns six isolated frames and surfaces over 30 ideas across clusters like economic-incentive, async-control-surface, gamification, perceptual-distortion, collective-intelligence, and redundancy-race. The critic pass then flags 20 traps with one-line reasons. The key difference from a single tree-of-thought is the isolation: branches do not share context, so one branch's early commitment cannot bias another. The critic pass is separate, which allows scoring and pruning after divergence. This design is more than a prompt template; it is a pipeline of parallel generation followed by evaluation.
Getting It Running: One Command, Agent Auto-Detection
Installation is a single command: npx skills add UditAkhourii/adhd. The README states it auto-detects the agent, including Claude Code, Cursor, Antigravity, Codex, Cline, Gemini CLI, Windsurf, and roughly 50 more. After install, you invoke it explicitly with /ad... (the README is truncated, but the pattern is a slash command). The project targets Node.js 18 or later, per the badge. The package is on npm as adhd-agent. No manual configuration keys are shown in the material, which suggests a zero-config setup for supported agents.
Evidence and Claims: What the README Actually Shows
The README includes one side-by-side eval problem with full transcripts in bench/results.json. An independent LLM judge scored breadth 9 versus 6, novelty 8 versus 3, and trap detection about 8 versus 2 for ADHD versus baseline. The methodology is in documentation/evals.md. The README also lists 17+ projects that ship or integrate ADHD, including repowire, mstack, and han. A research review with 11 sources and 8 validation rounds is cited, with findings tracked as issues #16 to #18. These are concrete artifacts, but they are not peer-reviewed benchmarks. The preprint link is to adhdstack.github.io, but the paper itself is not described in detail.
A Genuine Limitation: Cost and Overhead for Simple Tasks
The mechanism is expensive by design. Spawning multiple isolated reasoning processes and running a separate critic pass consumes significantly more tokens and latency than a single chain-of-thought call. For a well-specified coding task like 'add a type annotation' or 'fix this syntax error', this overhead is wasteful. The README itself positions the skill for divergent tasks, not routine ones. The eval shows ADHD surfacing 30+ ideas and 20 traps for a single problem, which is overkill when a senior engineer would give a textbook answer in 30 seconds. The skill is the wrong tool when you need a fast, deterministic answer and do not need to explore a design space.
Real Alternative: Standard Tree-of-Thought or Chain-of-Thought
The README contrasts ADHD with linear chain-of-thought and standard tree-of-thought. Standard tree-of-thought widens the search but walks a single shared context, so anchoring persists across branches. ADHD's difference is isolation: it spawns N processes with zero shared context during divergence. A practical alternative is to use a standard tree-of-thought library or even a simple prompt that asks for multiple approaches in one context. That approach is cheaper and easier to implement, but it suffers from the shared-context anchoring that ADHD targets. If you are on Claude Code, you could also use the built-in /think or /plan modes, but those do not provide the same parallel isolation.
Maintenance, License, and Adoption Signals
The project is MIT licensed, which is permissive for commercial use. The last push was August 2026, and the latest release v0.1.4 came in May 2026, indicating active maintenance. The README mentions a community is open, with a Discord and a contributor form. The first OSS adopter is repowire, which ported ADHD onto its mesh-orchestrator primitives in PR #313. The project has badges for CI and npm, but no explicit documentation on upgrade paths or versioning policy. As a skill, updates are likely distributed via the same npx skills add command, but the material does not detail that. The maintenance cost is low for users because the skill is a package, but the underlying agent SDKs may change, requiring updates.
Verdict: Who Should Adopt and What to Verify First
ADHD is worth trying for engineers who use coding agents for open-ended, interdisciplinary problems where premature convergence is a real risk. The side-by-side eval demonstrates a concrete benefit: the non-obvious pick of switching to a cheaper model was something baseline never considered. That kind of insight is valuable in design and strategy work. For routine coding, skip it. Before adoption, verify that your specific agent is in the auto-detection list, and run the eval problem from bench/results.json yourself to see if the breadth and novelty gains reproduce with your model and prompt. Check the preprint and the research review issues to understand the evidence base. The project is young, but the architectural approach is distinct and worth a controlled trial.
Editorial conclusion
ADHD is for engineers who regularly ask agents open-ended questions like design, naming, strategy, or fuzzy debugging and who are willing to pay extra tokens for breadth and trap detection. It is not for simple, well-specified coding tasks where a single pass is cheaper and faster. Before adopting, verify that your agent supports the skill auto-detection or manual install, and check the eval methodology in documentation/evals.md to see if the reported breadth and novelty gains hold for your workload. The project is MIT licensed and actively maintained, but the preprint and community adoptions are early signals, not proof of universal benefit.
Community notes