> ## Documentation Index
> Fetch the complete documentation index at: https://adhd.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.
# Introduction
> ADHD is a skill for coding agents — an architectural fix for premature convergence in autoregressive reasoning.
## What is ADHD?
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 — a hardware engineer, a regulator, a 10-year-old, a speedrunner — 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…"*.
Run your first ADHD loop as a skill, from the CLI, or in TypeScript.
The two-phase diverge/focus loop and the isolation invariant that makes it work.
The 15 cognitive frames, how selection works, and how to author your own.
Head-to-head numbers against a single-shot baseline, judged by an independent LLM.
## Why not just prompt harder?
The first three answers a model gives are the answers a senior engineer would give in thirty seconds. Correct. Forgettable. The interesting answers live past number three, in the awkward middle nobody walks into.
A single chain told to "consider alternatives" generates those alternatives **into one shared context, sequentially** — it anchors on whichever alternative it produces first and every subsequent "alternative" drifts toward it. ADHD's branches never share a context during divergence, so there is no first answer to anchor on. Distinctness is produced **by construction**, not requested by a prompt and hoped for.
> *"List alternatives" varies the **output**; ADHD varies the **generator**. Only the second escapes the anchor.*
## The loop in one picture
```text theme={null}
┌──────────────────────────┐
│ Problem P │
└────────────┬─────────────┘
│ pick N frames
┌──────────┬──────────┬──────┴───┬──────────┬──────────┐
▼ ▼ ▼ ▼ ▼ ▼
[hardware] [regulator] [10-y-old] [biology] [speedrun] … isolated,
branch branch branch branch branch parallel,
│ │ │ │ │ no critic
└──────────┴──────────┴────┬─────┴──────────┘
▼
┌──────────────────────────┐
│ Critic (separate call) │
│ score → cluster → prune │
│ traps → deepen top-K │
└────────────┬─────────────┘
▼
wide set · shortlist · ★ non-obvious pick
traps with reasons · deepened sketches · provocation
```
1. **Diverge.** N parallel, **isolated** LLM 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.
2. **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.
The generator–critic split is **mechanical** — separate LLM calls with opposite system prompts — not promised in one prompt.
## Baseline vs ADHD, one problem
> **Problem.** *"We have a CLI that calls an LLM and it sometimes hangs for 90 seconds. Design the right retry/timeout/UX strategy."*
Walks through four **textbook** patterns: progressive timeout with staged UI, fast-fail + exponential backoff, hedged parallel requests, streaming with keepalive. Lands on a sensible hybrid recommendation — 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 and surfaces a **wide set** of 30+ ideas across economic-incentive, async-control-surface, gamification, perceptual-distortion, collective-intelligence, and redundancy-race clusters, then:
* ★ **Non-obvious pick:** *"rage-quit = instant abort + branch to cheaper/faster model"* — one click cancels and re-submits to a Haiku-class model. The thing baseline never considers: **the slow model might just be the wrong model for this prompt.**
* Plus a shortlist: scout-fork to alternate endpoints at 30s; daemonize the CLI with ticket IDs; race 3 LLM replicas and cache the winner.
* **20 traps flagged with one-line reasons** — before they cost engineering time.
Independent LLM judge on this problem: **breadth 9 vs 6, novelty 8 vs 3, trap detection \~8 vs \~2**. See [Evals](/evals/results) for the full suite.
## Three ways to run it
`npx skills add UditAkhourii/adhd` — works in Claude Code, Cursor, Codex, and \~50 more agents. No API key needed.
`npm install -g adhd-agent` — run the loop from any terminal, with structured JSON output.
`npm install adhd-agent` — call `run()` at decision points inside your own agent.
## Project
ADHD operationalizes the *Divergent Ideation* source spec and ships under the MIT license.
* 📄 **Preprint:** [ADHD: Parallel Divergent Ideation for Coding Agents](https://adhdstack.github.io/)
* 📰 **Featured:** [The New Stack ran a feature story](https://thenewstack.io/claude-code-adhd/) on ADHD for Claude Code
* 👤 **Author:** Udit Akhouri — [@akhouriudit](https://x.com/akhouriudit) · [GitHub](https://github.com/UditAkhourii)
* 💬 **Community:** [Discord](https://discord.gg/NbWwkwwGw) · [join as a contributor or adopter](https://tally.so/r/WO1Nzj)