CLI Reference - ADHD

Install globally and authenticate with ANTHROPIC_API_KEY (or inherit auth from a local Claude Code install):

npm install -g adhd-agent

Examples

adhd "design a rate limiter that survives a leader election"

adhd "name this function" --frames 3 --ideas 8 --top 2

adhd "we have a CLI that hangs for 90s on LLM calls. what's the right retry/UX?" \
    --frames 5 --ideas 6 --top 3 --context ./client.ts

adhd "..." --json > result.json

Flags

Flag	Default	What it does
`--frames N`	5	parallel divergence branches
`--ideas N`	6	ideas per branch
`--top N`	3	how many ideas to deepen / focus
`--concurrency N`	4	max parallel LLM calls
`--context PATH`	—	inject a file as context (code, stack, constraints)
`--model NAME`	SDK default	override model (generator + critic)
`--critic-model NAME`	= `--model`	override model for the critic passes only (score + cluster) — use a different family to decorrelate critic errors
`--no-code-mode`	—	don’t bias frames toward engineering
`--no-anchor-strip`	—	don’t strip incidental anchors (stack, tool names) from the problem before fan-out
`--json`	—	emit machine-readable `RunResult`
`--quiet`	—	suppress progress events

--critic-model exists so the judge and the generator don’t share failure modes: scoring with a different model family decorrelates critic errors from generator errors.

Output

The default flow prints, in order: the clustered wide set with score chips, a 2–4 idea shortlist with the ★ non-obvious pick, the trap list with one-line reasons, the top-K deepened sketches, and a provocation. With --json, the same content is emitted as a structured RunResult for downstream programs.

⌘I