I-have-ADHD: A skill to stop coding agents from burying the answer

Original Article ↗ Hacker News Discussion ↗

LLM users are turning to a GitHub “i-have-adhd” skill and similar prompt hacks to force coding agents—especially Claude/Opus 5— to stop burying answers in verbose, TED‑talk‑style prose and to lead with concise, actionable output. Commenters debate whether this should be a per‑user/output setting rather than a separate skill, how well such instructions persist over long sessions, and whether Anthropic’s training and watermarking choices have structurally pushed its models toward unreadable slop. The ADHD framing itself is controversial, with some diagnosed users finding the name trivializing even as many note that “ADHD‑friendly” formats—BLUF, short lists, concrete next steps—tend to improve usability for almost everyone.

Purpose of the “I-have-ADHD” Skill

  • Skill is meant to stop coding agents (especially Claude) from burying answers in long, meandering prose.
  • Emphasizes: concise output, BLUF (bottom line up front), concrete next steps, and ADHD‑friendly formatting.
  • Several commenters note they already use similar tricks: “I have ADHD, be concise”, “ELI5”, “executive on a phone”, “grug brain”, etc.

Effectiveness and Limitations

  • Some users report a “night and day” improvement, especially when the skill is re‑asserted every turn or used as a stop‑hook.
  • Others say Claude follows it for only a few exchanges before reverting to verbosity.
  • A few say it “doesn’t work” or only marginally helps, especially with Opus 5.
  • Consensus: style instructions decay over long sessions; skills alone cannot fully fix model behavior.

Claude’s Verbosity and “Claudisms”

  • Strong frustration with Opus 5’s TED‑talk style: over-explaining, hedging, nitpicking, and burying the lede.
  • Common complaints:
    • Long background essays for simple questions.
    • Reporting what it did not do, and documenting abandoned approaches.
    • Load-bearing clichés, self-narration, and SEO‑like filler.
  • Many downgrade to earlier Claude versions (4.6/4.8) or switch to other providers.

Alternatives: Output Styles, Instructions, Other Models

  • Some prefer custom output styles or global CLAUDE.md rules (e.g., ASD‑STE100, “tersely”, “no em-dashes”, “no historical remarks”).
  • Mixed reports on Claude Code’s “Concise” output style: helps somewhat, but doesn’t fully stop yapping.
  • Several say other models (Codex/Astra/GPT variants) are more concise and follow style instructions better.

ADHD Framing and Neurodiversity Concerns

  • Mixed reactions from people with diagnosed ADHD:
    • Some like the framing and say ADHD-informed formatting helps everyone.
    • Others find the branding cringe or trivializing, noting ADHD is not just “not wanting to read walls of text.”
  • There’s concern that casual “I have ADHD” prompts may dilute accommodations for people who truly need them.

Skills, Repo Design, and Safety

  • Debate over why this is a “skill” vs just AGENTS.md / output style; some see it as fashion or distribution convenience.
  • Questions about the large repo size; much of it is evals and harness-specific glue.
  • A few are uneasy about copy‑pasting “install” commands that let agents fetch and run code from GitHub.