i-have-adhd skill stops coding agents from burying answers

ayghri ships i-have-adhd, a skill that reshapes agent output

Developer ayghri released i-have-adhd, an MIT-licensed skill that changes how coding agents respond to a prompt. Instead of a preamble, a recap and a closing "hope this helps," the skill forces the agent to lead with the next action, number every multi-step task and end with one concrete thing to do next. Posted to Hacker News on 8 September 2026, the story reached 347 points and 266 comments.

i-have-adhd is a SKILL.md file that coding agents load to change their output shape rather than their reasoning: command or snippet first, numbered steps for multi-step work, one concrete next action at the end, no preamble and no closing pleasantries. It installs into Claude Code, Codex, Gemini CLI, GitHub Copilot, OpenCode, Antigravity, Hermes and Kimi Code CLI.

What changed

• The skill sets out ten rules in full in skills/i-have-adhd/SKILL.md: lead with the next action, number multi-step tasks, end with one concrete next action, suppress tangents, restate state every turn, give specific time estimates, make completed work visible, use a matter-of-fact tone for errors, cap lists at five items, and cut preamble, recaps and closing pleasantries.
• The README's own before/after example shows the shift: a paragraph that opens "Great question! Let me think about this..." becomes "Run npm install jsonwebtoken@latest, then edit src/auth.ts:42" followed by three numbered steps and a named next action.
• The SKILL.md frontmatter sets disable-model-invocation: true, so no harness auto-loads the skill. A user invokes it with a slash command (/i-have-adhd in Claude Code and OpenCode, $i-have-adhd in Codex) or wires an always-on route: a ~/.claude/.i-have-adhd-always flag file that fires a Claude Code SessionStart hook, a block pasted into ~/.gemini/GEMINI.md, .github/copilot-instructions.md, AGENTS.md or a persona SOUL.md.
• GitHub Copilot reads Agent Skills natively from the same SKILL.md, no conversion needed, scanning .github/skills/, .claude/skills/ and .agents/skills/ in a project and the equivalent home-directory folders globally; the maintainers note Copilot also respects disable-model-invocation, tested in pull request #60.
• The rules are credited as loosely based on The Adult ADHD Tool Kit by J. Russell Ramsay and Anthony L. Rostain, adapted for how an LLM should respond rather than how a human should organize a day.

Why it matters for agent builders

The repository packages one SKILL.md and ships eight separate install paths (Claude Code plugin marketplace, Codex plugin, a Gemini CLI command or extension, a Copilot skills folder, an OpenCode server plugin, an Antigravity plugin, a Hermes skill tap, a Kimi Code CLI custom plugin), which is a working example of how far one skill file now travels across harnesses without a rewrite. disable-model-invocation is a concrete, checkable control surface: anyone packaging a skill can decide whether it activates automatically or waits for an explicit call, and this skill chooses the latter by default.

The Hacker News thread also surfaces a specific limit worth knowing before writing custom rules for an agent. User ryandrake described Claude Code's habit of narrating what it did not do ("I edited this.py but did not edit README.md"). Commenter hysan reported that across more than 30 sessions of asking the model to explain its own contradictions, it consistently blamed harness-level instructions rather than its own training or a user's custom rule, concluding that when a custom rule conflicts with a harness instruction, the harness tends to win.

What to watch

The thread splits on whether that harness-versus-model explanation is even testable, since a model narrating its own reasoning cannot be trusted to accurately report the cause of its own behavior; one reply called the model's self-diagnosis "something plausible sounding" rather than a real memory of what happened. Whether i-have-adhd's ten rules hold up against that same harness-level override, across more sessions and more harnesses, is the open question the repository's issue tracker will settle over time.

Key takeaways

• i-have-adhd is a single SKILL.md file, MIT-licensed, installable across eight different coding-agent harnesses with harness-specific commands.
• It is off by default: disable-model-invocation: true means a user must invoke it with a slash command or wire an always-on hook or config block.
• The Hacker News discussion (347 points, 266 comments) surfaced a harness-versus-model debate over why agents keep narrating what they did not do, with one commenter's 30-plus-session test pointing at the harness.
• GitHub Copilot loads the same SKILL.md with no conversion step, reading it from a set of local and global skills folders.

FAQ

What is i-have-adhd?

It is an open-source skill for coding agents that changes output formatting: action first, numbered steps for multi-step work, one concrete next action at the end, and no preamble, recap or closing pleasantry. It does not change what the agent reasons about, only how it presents the result.

How do I turn on i-have-adhd in Claude Code?

Install it with claude plugin marketplace add ayghri/i-have-adhd followed by claude plugin install i-have-adhd@i-have-adhd, then type /i-have-adhd in a session. For it to load automatically at every session start instead, create the file ~/.claude/.i-have-adhd-always, which triggers a SessionStart hook the plugin ships.

Does i-have-adhd run automatically once installed?

No. Its SKILL.md sets disable-model-invocation: true, so installing the plugin changes nothing by itself; the rules apply only once a user invokes the skill explicitly or sets up one of the always-on routes documented for each harness.

Source: i-have-adhd on GitHub Discussion: Hacker News

Developers comparing packaged skills like this one against other agent tooling can browse the full list on aiagentslisting.com's skills directory.