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.
🎮 Join the Discord → for frame design, eval problems, and trap-hunting in real time · 👉 Join the community → as a contributor, maintainer, or early adopter (one short form).
An architectural fix for premature convergence in autoregressive reasoning.
Linear Chain-of-Thought anchors on whatever it 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 an architectural problem, not a prompting one — 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.
Reach for it on design decisions, fuzzy debugging, naming, API surface design, strategy, and any prompt of the shape “give me a few ways to…”.
One eval problem, same model, two strategies. Full transcripts in bench/results.json.
Problem.“We have a CLI that calls an LLM and it sometimes hangs for 90 seconds. Design the right retry/timeout/UX strategy.”
Baseline gives the sensible textbook hybrid (staged timeouts + one auto-retry) — the answer a senior engineer gives in 30 seconds, with no traps named. ADHD spawns 6 isolated frames, surfaces 30+ ideas, flags 20 traps with reasons, and lands the non-obvious pick baseline never considers: the slow model might just be the wrong model for this prompt — instant abort + branch to a cheaper/faster one.
Expand the full side-by-side
🟦 Baseline (single-shot)
🟧 ADHD
Walks through four textbook patterns:
Progressive timeout with staged UI (10s / 30s / 60s)
Fast-fail + exponential backoff retry
Hedged parallel requests
Streaming with keepalive
Lands on a hybrid recommendation — 15s first-token timeout, 30s between-token timeout, 90s absolute, one auto-retry. Sensible. Google SRE Book ch. 22. The answer a senior engineer gives in 30 seconds.
What’s missing: no traps named, no acknowledgement that the user might want to bail out of a slow request, no questioning of the “wait then retry the same model” frame.
Spawns 6 isolated frames, surfaces a wide set of 30+ ideas across economic-incentive, async-control-surface, gamification, perceptual-distortion, collective-intelligence, redundancy-race clusters, then:
★ Non-obvious pick:“rage-quit = instant abort + branch to cheaper/faster model” — a button that pulses hotter the longer you wait. One click cancels and re-submits to Haiku-class. The thing baseline never considers: the slow model might just be the wrong model for this prompt.
Plus shortlist: scout-fork to alternate endpoints at 30s; daemonize the CLI with ticket IDs; race 3 LLM replicas, cache the winner.
20 traps flagged with one-line reasons — including the cute “stream tokens in reverse” and “patience-token billing” ideas before they cost engineering time.
Independent LLM judge on this problem: breadth 9 vs 6, novelty 8 vs 3, trap detection ~8 vs ~2. Methodology in documentation/evals.md.
Featured
🔌 Adopted by repowire — the first OSS project to officially ship ADHD, ported onto its mesh-orchestrator primitives in PR #313 (merged).
📰 The New Stack ran a feature story on ADHD for Claude Code.
💬 OpenClaw / multi-agent community is independently testing it across agents. One tester: “I read it, installed it on two different agents… I actually love it. This is great. I thought this was gonna be another useless post. But no, it wasn’t.”
A two-phase loop with a hard wall between the phases.
Diverge. Pick N cognitive frames. Spawn N parallel, isolated Agent calls — each sees the problem plus one frame’s vantage prompt, and a system prompt that forbids evaluation. Branches never see each other, so no anchoring.
Focus. A separate critic call scores every idea (novelty / viability / fit), flags traps with reasons, clusters by underlying angle, and deepens the top-K survivors into sketches with risks and first steps.
Mean scores across 6 open-ended engineering problems (0–10), ADHD vs a single-shot baseline at the same model, judged by an independent LLM with a skeptical-staff-engineer prompt, A/B order randomized.
Dimension
ADHD
Baseline
Δ
Ratio
breadth
9.00
4.83
+4.17
1.9×
novelty
7.83
2.67
+5.17
2.9×
trap detection
9.50
1.83
+7.67
5.2×
actionability
9.50
6.50
+3.00
1.5×
builder usefulness
7.67
6.83
+0.83
1.1×
ADHD wins 5 of 6 problems. Biggest gap is trap detection — baselines rarely name the seductive-but-broken ideas. Methodology, limitations, and how to reproduce: documentation/evals.md.
Documentation
📚 Official docs:adhd.mintlify.site — the full, browsable documentation site.
A measured duel vs. single-shot by Shichinomiya (@shichinomiya_s) — independent blind-scored benchmark (2 problems, LLM-as-judge, A/B positions swapped). ADHD won both, with the biggest gains in novelty (4.5→9.0) and trap detection (5.0→9.0), at a real cost of ~2.3× time and ~1.9× output.
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?v=4)
评论