i-have-adhd Skill for Claude Code: ADHD-Friendly Replies, Action First

• Turns the style on; in Claude Code it waits for you to call it

  Type it to start. The rules then apply to every reply for the rest of the session, even when the topic changes. In Codex the call is $i-have-adhd.

  Try: “/i-have-adhd”

  /i-have-adhd

• Turns the style off; "normal mode" works too

  Say either phrase. The agent confirms in one line and goes back to its default style.

  Try: “stop adhd mode”

  stop adhd mode

• Why the rules exist

  Working memory is small, knowing is not doing, starting is hard, time estimates feel uniform, and visible progress matters because dopamine is scarce.

  Five facts about ADHD reading

• First line is something you can do now

  A command, path or snippet goes first and prose comes after, if at all.

  Rule 1: Lead with the next action

• One bounded action per step, fewest steps that work

  Cut steps you do not need and fold trivial ones in. A short path finished beats a complete path abandoned.

  Rule 2: Number multi-step tasks

• One thing you can do in under two minutes

  If anything is left open, the reply names a single next step, even if it is just opening a file.

  Rule 3: End with one next action

• Finish the first issue, then offer the second as a question

  A question that comes up mid-work is folded in if the agent can answer it. If it still needs you, it is raised once at the end.

  Rule 4: Suppress tangents

• Say which step you are on, for example step 3 of 5

  Uses the harness's task or plan tool for multi-step work, so the checklist does the restating.

  Rule 5: Restate state every turn

• Minutes or an afternoon, not "a bit of work"

  Example: about 15 minutes if tests already cover this, an afternoon if not.

  Rule 6: Give specific time estimates

• Say what now works and how to try it

  Show the win in concrete terms instead of burying it in a recap.

  Rule 7: Make completed work visible

• State cause and fix; no "Uh oh"

  Errors name the file and line, the cause and the fix.

  Rule 8: Matter-of-fact tone for errors

• Group and rank; this shapes display only

  Extra items are kept and shown when you ask or when they come up next. The rule does not limit analysis, search or tool results.

  Rule 9: Cap lists to 5 items

• Start with the answer and end when it is done

  Banned openers include "Great question" and "Let me...". Banned closers include "Hope this helps".

  Rule 10: No preamble, recap or closer

• Six cases, such as explain requests and destructive actions

  Explain requests run long. Destructive actions need confirmation. After three failed fixes it names the doubtful assumption. Real ambiguity gets one question. The task or the harness can override a rule.

  When to break the rules

• Deletes announcement openers, recaps, sidebars and filler hedges

  Then checks that the first and last lines alone tell you what to do next and what just happened.

  Pre-send check

• Creating ~/.claude/.i-have-adhd-always loads the rules every session

  Optional, from INSTALL.md. A SessionStart hook reads the flag file. Delete the file to go back to on-demand.

  Try: “touch ~/.claude/.i-have-adhd-always”

  Runs code · Always-on flag (plugin)

Watch out for

• The plugin adds a hook that runs when a session starts

  hooks.json runs a node command on startup, resume, clear and compact. It looks for ~/.claude/.i-have-adhd-always (or the same file in CLAUDE_CONFIG_DIR) and only if that file exists prints the SKILL.md rules into the session. Any failure exits quietly. github.com

• The style stays on for the whole session

  SKILL.md says the rules do not lapse when the topic changes. Say "stop adhd mode" or "normal mode" to turn them off. github.com

• Destructive actions still get a confirmation

  SKILL.md lists exceptions: it confirms before things like rm -rf, force push or dropping a table, and explains fully when you ask it to explain. github.com