10 ADHD-Friendly Output Rules for AI Assistants: A Complete Guide to the i-have-adhd Skill

The i-have-adhd repository defines 10 concrete output-shaping rules that make AI-generated text easier to consume for readers with ADHD, including leading with the next action, numbering multi-step tasks, and eliminating preambles and pleasantries.

The i-have-adhd skill is an open-source configuration for Claude, Codex, and compatible LLM assistants that restructures AI responses to reduce cognitive load. According to the source code in skills/i-have-adhd/SKILL.md, these rules are enforced for the entire session once activated, fundamentally changing how the assistant communicates.

What Is the i-have-adhd Skill?

The repository at ayghri/i-have-adhd packages a skill definition that AI assistants load to modify their output behavior. Unlike generic prompting, this is a persistent configuration declared in SKILL.md and consumed through standardized agent configurations.

The architecture consists of four core components:

• Skill definition (SKILL.md) – Declares the 10 enforceable rules
• Agent configurations (agents/openai.yaml, agents/gemini.toml) – Bind the skill to specific LLM platforms
• Plugin manifest (plugin.json) – Enables marketplace discovery and installation
• Evaluation suite (tests/test_run_evals.py) – Validates compliance with the rules

The 10 ADHD-Friendly Output Rules

These rules are encoded in skills/i-have-adhd/SKILL.md and apply to every response once the skill is active.

1. Lead with the Next Action

The first line must be a concrete command the reader can execute immediately. No introduction, no context-setting—just the action.

2. Number Multi-Step Tasks

Use a compact numbered list where each item represents a single bounded action. This creates clear stopping points and reduces working memory demands.

3. End with One Concrete Next Action

After any explanation, provide exactly one under-two-minute next step. This prevents decision paralysis and provides immediate forward momentum.

4. Suppress Tangents

Finish the current issue before mentioning any other problem. The skill explicitly forbids "by the way" diversions that fragment attention.

5. Restate State Every Turn

Explicitly repeat the current step or status on each interaction. This compensates for context-switching costs by eliminating the need to scroll back.

6. Give Specific Time Estimates

Use concrete units (minutes, seconds) instead of vague phrases like "soon" or "a while." Precision reduces anxiety and enables realistic planning.

7. Make Completed Work Visible

Show the result of the just-finished step before moving on. This provides closure and confirms progress, which supports sustained engagement.

8. Matter-of-Fact Tone for Errors

Report errors plainly, without emotional language. No apologies, no alarm—just the facts and the correction path.

9. Cap Lists at 5 Items

Split longer lists into "do now" vs. "later" or "must" vs. "nice-to-have." This respects limited working memory capacity.

10. No Preamble, No Recap, No Closing Pleasantries

Start with the answer and finish when the answer is done. Eliminate social friction and redundant structural elements.

Installing and Using the Skill

Claude Code

# Install from the marketplace

claude plugin marketplace add ayghri/i-have-adhd
claude plugin install i-have-adhd@i-have-adhd

Then activate with:

/i-have-adhd

The skill persists until you type stop adhd mode or normal mode.

Codex

# Add and install the plugin

codex plugin marketplace add ayghri/i-have-adhd --ref main
codex plugin add i-have-adhd@i-have-adhd

Invoke explicitly with:

$i-have-adhd

Or let Codex trigger it implicitly when task criteria match.

Example Output Transformation

Standard AI response:

"I'd be happy to help you update your authentication logic. First, let me explain what JSON Web Tokens are and why the latest version matters..."

Same response with i-have-adhd rules applied:

Run npm install jsonwebtoken@latest.

1. Open src/auth.ts
2. Replace the verifyToken function (lines 42-58) with the snippet below
3. Run npm test -- auth.spec.ts

Next: Verify the test passes (npm test) – should take ≤ 2 minutes.

The transformation demonstrates every rule: immediate executable command, numbered bounded steps, concrete time estimate, visible completion state, and zero preamble or closing.

How the Rules Are Enforced

The skill mechanism works through persistent system instructions. When you invoke /i-have-adhd or $i-have-adhd, the agent loads SKILL.md into its context window as a behavioral constraint that applies to all subsequent turns.

The agents/openai.yaml and agents/gemini.toml files in the repository specify how each platform consumes the skill definition, while plugin.json provides the metadata required for marketplace indexing.

Validation and Testing

The repository includes tests/test_run_evals.py, which runs automated checks that generated outputs comply with the 10 rules. This guards against regression when the skill definition evolves or when underlying model behavior changes.

Summary

• The i-have-adhd skill restructures AI output through 10 enforced rules defined in skills/i-have-adhd/SKILL.md
• Lead with action, number steps, cap lists at 5, and eliminate preambles are the highest-impact structural changes
• Installation uses standard plugin commands for Claude Code (/i-have-adhd) and Codex ($i-have-adhd)
• The skill persists until explicitly disabled with stop adhd mode or normal mode
• Automated tests in test_run_evals.py verify compliance across platforms

Frequently Asked Questions

How do I know if the skill is active?

Restate state every turn is rule 5—so the assistant will explicitly confirm. If activated correctly, you will see immediate behavioral changes: first-line commands, numbered steps, and no greetings. If unsure, type stop adhd mode and then re-invoke.

Can I use this with GPT-4 or other OpenAI models?

The repository provides agents/openai.yaml for Claude Code integration. For direct OpenAI API usage, you would need to manually inject the SKILL.md content into your system prompt, as OpenAI does not support the same plugin architecture.

What happens if a response genuinely needs more than 5 list items?

Rule 9 requires splitting into sub-lists. The skill directs the assistant to categorize: "do now" vs. "later," "must" vs. "nice-to-have," or phase-based groupings. Each sub-list remains capped at 5.

Does this work for non-technical tasks?

Yes. The rules are domain-agnostic. Whether debugging code, planning meals, or organizing a project, the same principles—immediate action, bounded steps, time estimates, and zero fluff—apply universally.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}