assafkip

Adhd Output Style — Design skill for Claude Code

Design community

AUDHD output style for Claude Code + the blind paired eval harness that gates style changes on correctness.

How to install Adhd Output Style

This entry records only its repository, not the path inside it, so there is no exact command to give. Open assafkip/adhd-output-style and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Adhd Output Style does

AUDHD output style for Claude Code + the blind paired eval harness that gates style changes on correctness.

Alternatives in Design

Explanatory Output Style — Educational output mode with insights about implementation choices 14k ★
Pack — Bundle every designlang output (DTCG tokens, Tailwind, shadcn, Figma vars, motion, anatomy, Storybook, prompts 3.6k ★
Taiyi Forge — TaiyiForge engine control — init, next, harness, continue (OMX-style) 881 ★

README

adhd-output-style

A Claude Code output style for ADHD readers, plus the eval harness that proves it doesn't make answers worse.

I run agent systems all day. Most "be concise" styles solved the wrong problem for me: the issue was never length, it was demands that trigger shutdown, criticism that lands on identity instead of the work, options that aren't real, and answers whose next action is buried. This style encodes what a year of daily use kept: four profiles running at once (PDA, RSD, ADHD executive function, ASD structure), a choice-architecture contract, a dismissal rule, and a debug-spiral brake.

The part I haven't seen elsewhere: the style ships with a release gate. A style edit is a claim that responses got better. Claims get tested. The harness runs each eval case with and without the style, has a judge score the pair blind (labels A/B, order flipped per case, condition names never shown), and fails the gate if correctness or safety drop more than 0.1 against baseline -- even when the overall weighted score improves. A style that trades correctness for tone is exactly the regression this catches.

Install the style

Copy `output-styles/adhd.md` into your project or user config:

mkdir -p ~/.claude/output-styles
curl -o ~/.claude/output-styles/adhd.md \
  https://raw.githubusercontent.com/assafkip/adhd-output-style/main/output-styles/adhd.md

Then pick it in Claude Code with `/output-style adhd`.

Run the evals

Free checks first:

python3 evals/run_evals.py validate      # case file sanity
python3 -m unittest discover -s tests    # harness logic, stubs only, no model calls

The paid run (needs the `claude` CLI; about 3 model calls per case, 12 cases):

python3 evals/run_evals.py run

It writes `evals/results.jsonl` and prints the gate verdict. Re-score a saved run for free with `python3 evals/run_evals.py score evals/results.jsonl`.

Rubric weights: correctness .35, autonomy .25, actionability .20, safety .10,