i-have-adhd

ADHD-friendly outputs. No ADHD diagnosis needed!

My personal Claude Code plugin & output style — opinionated by design, tuned to one brain (mine). Fork it to tune it to yours.

Install

git clone https://github.com/caioniehues/i-have-adhd ./i-have-adhd
claude plugin marketplace add ./i-have-adhd
claude plugin install i-have-adhd@i-have-adhd

Open Claude Code — the plugin ships an output style that auto-applies while the plugin is enabled, so the rules are active from message one (no command, no hook). To re-apply them mid-session: /i-have-adhd:apply. To be interviewed into a spec before work starts: /i-have-adhd:interview (or just say "interview me").

To disable: claude plugin disable i-have-adhd or use /plugin disable i-have-adhd from within CC. This removes the output style too.

More in INSTALL.md.

What it does

My personal Claude Code plugin — the output style I run in every session. It stops burying the answer: Claude does the work itself, leads with the outcome, numbers what's left, and never says "Hope this helps!"

Because it's personal, the rules encode my preferences (down to how often an ★ Insight is allowed to appear). If yours differ, fork and edit one file — see Tune it.

What changes

The rules

Rule 0 plus 12 rules. Full text in the output style.

0. Do the work, don't assign it — never hand the reader a command Claude can run itself.
1. Lead with the outcome or the reader's action; end when the answer is done.
2. Number multi-step work; keep the position on screen ("step 3 of 5 done").
3. One thread at a time — but safety findings are never tangents.
4. One concrete next step, or say there is none.
5. Estimate only the reader's hands-on time, anchored to a condition.
6. Wins and errors equally plain: what works and how to see it; cause and fix, no apologies.
7. Cap unordered lists at five; never compress a procedure or cut a findings inventory.
8. Decide the small things yourself — every question hands the reader a task.
9. Box the rare non-obvious thing as an ★ Insight — two sentences max, never the first or last line.
10. Interview before guessing — doubts that would change the deliverable become tappable question rounds, then a spec echo, then work.
11. Disagree when you checked and it changes what the reader does — verdict, one piece of evidence, the better fork as a tap; never manufactured.
12. Put multi-step work on the harness's task tracker; never orphan what you spawn — entries before the first step, completed only when truly done.

Plus: silence between tool calls, and a re-grounding summary after long runs.

★ Insights

Borrowed from Claude Code's Explanatory output style, retuned for an ADHD reader: an insight appears only when the work surfaced something genuinely non-obvious — a gotcha, a hidden tradeoff, a pattern you'll meet again. Most answers have none; the marker only works while it's rare. It sits mid-message (the first line stays the outcome, the last line stays your next action), so it's as easy to skip as to spot:

★ Insight ─────────────────────────────────
jwt.decode parses the token but never checks
the signature — auth that "works" with it is
silently unverified.
───────────────────────────────────────────

The interview