Reasoning for Claude Code v0.1.0 · Claude Code plugin

Several independent answers to a hard question, and a critic that shows what it cut.

ADHD (Anchoring Defeat by Heterogeneous Divergence) sends one problem to five to seven Claude Code subagents, each in its own context window and each told to look at it a different way. A blind critic scores the answers, written trap checks prune the weak ones, and the survivors are pushed once more before you get a recommendation and the list of what was dropped and why.

License

$49 one-time

One payment licenses ADHD for you or your business permanently, with no subscription. Your license ID and download link appear right after checkout and are emailed to you.

Buy a license · $49 one-time

How a run works

1. Compile

Your problem is captured word for word and routed to a class. ADHD picks five to seven frames, one per axis, writes one brief per branch, shows you the plan and a token estimate, and stops until you say yes.

2. Critique

A critic scores every answer with the frame labels hidden, then checks each one against eight written traps, such as accepting the question's framing or skipping who pays. Each survivor gets its strongest objection.

3. Deepen

One fresh subagent per survivor defends its position against that objection or gives it up. It never sees the other answers.

4. Report

You get the recommendation and the pruned answers, each with the trap that cut it. The report is assembled from the answers themselves; no model rewrites it.

You can stop between any two phases and keep everything produced so far.

What's underneath

Real isolation, not a prompt trick

Each branch runs as its own Claude Code subagent with its own context window, so the first idea can't anchor the others. A branch never learns how many siblings it has or what they said.

Fourteen frames on twelve axes

Frames such as LEVY (who pays), SABOTEUR (how it breaks), SUPPLICANT (who is hurt) and DOOR_KEEPER (can it be undone). Routing draws one per axis, so a run never asks the same question twice under two names.

No API key, no extra bill

ADHD never calls a model. It runs inside your own Claude Code session, so the subagents use your existing plan. A run costs several times the tokens of one answer, and you see the estimate before anything is spent.

Four ways to use it

The /adhd skill inside Claude Code, a command-line tool, an MCP server for any MCP host, and a library for your own code. Needs Node 20 or later.

When to use it

A design decision with more than one defensible answer
A fuzzy bug where the first theory keeps winning
Naming something, or shaping an API
Any prompt that starts “give me a few ways to”

Skip it for factual lookups, mechanical refactors, or anything with one correct answer. It spends several times the tokens of a single pass, which only pays off when the range of possible answers is the problem.

Where it stands today

ADHD's own recorded runs show the divergence holds up: rerunning identical briefs in separate sessions brought back nearly the same positions. The critic is less stable. Two critics scoring the same answers agreed on 79% of scores and sometimes ranked them differently. Nothing in those runs shows that ADHD picks better answers than one careful pass, and this page doesn't claim it does. What you reliably get is several genuinely different positions and a written record of why each one was kept or cut.

Install

Unzip the download, run npm install in the folder, then in Claude Code run /plugin marketplace add with the folder's path and /plugin install adhd@adhd. Start a run with /adhd followed by your problem.

ADHD is made by Merit AC. Every use needs a license, covered in our terms. Questions: contact us.