How the i-have-adhd Plugin Changes AI Assistant Output: A Technical Deep Dive

The i-have-adhd plugin rewrites the conversation context by injecting ADHD-specific formatting rules into the system prompt, forcing the AI to produce structured, action-oriented responses with numbered steps and suppressed tangents.

The ayghri/i-have-adhd repository provides a Pi Coding Agent extension that fundamentally alters how an AI assistant generates responses. By intercepting the prompt pipeline and injecting custom behavioral rules, the i-have-adhd plugin transforms verbose, meandering outputs into concise, actionable guidance optimized for neurodivergent users.

Rule Injection via syncContext()

The primary mechanism for changing AI output occurs in extensions/i-have-adhd.ts within the syncContext() function (lines 122-136). When ADHD mode is enabled, the plugin inserts a custom message containing the ADHD-style rules from skills/i-have-adhd/SKILL.md into the conversation context.

// The plugin sends the rules as a custom message type
pi.sendMessage({
  type: 'i-have-adhd-rules',
  content: skillMarkdown, // Contents of SKILL.md
});

This injection happens via pi.sendMessage, placing the formatting constraints directly into the model's context window. Because these rules persist as a custom message, they effectively act as persistent system instructions for every subsequent generation.

State Tracking and Persistence

The plugin maintains awareness of whether ADHD mode is active through several state handling functions defined in extensions/i-have-adhd.ts (lines 60-71). The getSavedState and setEnabled functions manage a custom session entry called i-have-adhd-state that persists across agent sessions.

Users can activate the mode through multiple entry points:

• CLI Flag: Starting the agent with --adhd sets pi.getFlag("adhd") to true
• Always-on Marker: Creating a file like .i-have-adhd-always in the agent root directory
• Session Storage: The custom session entry maintains state between restarts

UI Status Indicators

When ADHD mode is active, the plugin provides visual feedback through the updateStatus() function (lines 8-16). This updates ctx.ui.setStatus to display a small "● ADHD ON" badge, ensuring users always know when the special output formatting is in effect.

function updateStatus(enabled: boolean) {
  ctx.ui.setStatus(
    enabled ? '● ADHD ON' : '',
    enabled ? 'active' : 'inactive'
  );
}

Command and Input Handling

Users control the plugin through the /i-have-adhd [on|off] slash command and natural language triggers. The pi.registerCommand handler (lines 73-95) processes explicit toggle commands, while an input listener (lines 97-110) intercepts phrases like "stop adhd mode".

When disabling the mode, the plugin sends a custom message of type i-have-adhd-disabled that removes the behavioral rules from the context, immediately restoring the model's default conversational style.

How the Injected Rules Reshape AI Output

The markdown in skills/i-have-adhd/SKILL.md defines 10 concrete style rules that constrain the model's generation behavior. Because these rules are passed as a persistent custom message, every response is biased to:

• Lead with the next action rather than explanations or pleasantries
• Number multi-step tasks using short, scannable lists
• Suppress tangents and eliminate preambles like "Certainly!" or "Here's what you need to know"
• Provide concrete time estimates for tasks rather than vague durations
• End with a single next-step cue that clearly indicates what the user should do next

The rules effectively function as a system prompt overlay, forcing the AI to prioritize executive function support over conversational fluency.

Implementation Details and Code Examples

Enable ADHD mode programmatically or via CLI:

// Check if enabled via CLI flag
if (pi.getFlag("adhd") === true) {
  console.log('ADHD mode active via --adhd flag');
}

// Toggle via slash command
await pi.executeCommand("i-have-adhd", "on");   // Enables rules injection
await pi.executeCommand("i-have-adhd", "off");  // Removes rules from context

// Natural language disabling
await pi.receiveInput({ text: "stop adhd mode" }); 
// Returns: "ADHD mode disabled."

Key Files and Architecture

File	Role
extensions/i-have-adhd.ts	Core extension logic: loads rules, tracks state, injects/removes context, registers commands, updates UI status
skills/i-have-adhd/SKILL.md	Authoritative markdown containing the 10 ADHD-style formatting rules
plugin.json	Plugin manifest declaring name, version, and metadata
hooks/always-on.*	Helper scripts (e.g., always-on.sh) for auto-enabling mode at agent startup

Summary

• The i-have-adhd plugin operates as a Pi Coding Agent extension that modifies the conversation context, not the model weights or training.
• Rule injection via syncContext() inserts skills/i-have-adhd/SKILL.md content as a persistent custom message, effectively altering the system prompt.
• State persistence across sessions is handled through getSavedState and setEnabled functions managing the i-have-adhd-state entry.
• Users control the mode through CLI flags (--adhd), slash commands (/i-have-adhd), or natural language ("stop adhd mode").
• When active, the plugin forces AI output to use numbered lists, concrete time estimates, and single next-step cues while suppressing conversational filler.

Frequently Asked Questions

How does the i-have-adhd plugin technically modify the AI's system prompt?

The plugin does not directly edit the system prompt object. Instead, it uses pi.sendMessage in the syncContext() function to inject a custom message of type i-have-adhd-rules containing the formatting constraints from SKILL.md. Because this message persists in the conversation context, the model treats these rules as binding instructions for all subsequent generations.

Can I toggle ADHD mode without restarting the agent?

Yes. The plugin supports runtime toggling through the /i-have-adhd [on|off] slash command or natural language phrases like "stop adhd mode" (handled in lines 97-110 of extensions/i-have-adhd.ts). When toggled off, the plugin sends an i-have-adhd-disabled message that removes the rules from the active context, immediately restoring default output behavior.

What specific formatting changes occur when ADHD mode is enabled?

According to the SKILL.md rules injected into the context, the AI must lead responses with the immediate next action, use short numbered lists for multi-step procedures, provide concrete time estimates instead of vague durations, end with a single clear next-step cue, and eliminate tangential chatter, preambles, and social pleasantries.

Where does the plugin store the user's ADHD mode preference?

The plugin persists state through a custom session entry named i-have-adhd-state, managed by the getSavedState and setEnabled functions (lines 60-71). Additionally, it checks for the --adhd CLI flag at startup and can detect always-on marker files in the agent root directory for automatic activation.

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