How Agent Harness Interact with i-have-adhd Skill Precedence

When the i-have-adhd skill operates inside an agent harness, the harness's system prompt overrides the skill's default confirmation behavior, forcing immediate tool execution while preserving ADHD-friendly formatting rules.

The ayghri/i-have-adhd repository defines a specialized skill for generating ADHD-accessible AI responses, but understanding how agent harness interact with i-have-adhd skill precedence is essential when deploying autonomous agents. When integrated into harnesses like OpenAI, Gemini, or Cursor, the skill's internal constraints yield to harness-level directives that control tool invocation and confirmation flows.

Skill Precedence Rules Defined in SKILL.md

The interaction between harness and skill is governed by explicit precedence rules documented in skills/i-have-adhd/SKILL.md. These rules determine which authority controls response behavior when conflicts arise between user formatting preferences and execution requirements.

Rule 6: When the Harness Overrides the Skill

Rule 6—"A rule fights the harness" establishes that inside an agent harness, the system prompt outranks the skill's constraints. According to lines 128-129 of SKILL.md, this means the skill must "announce a tool call when the harness requires it, do the work instead of asking 'want me to,' [and] point time estimates at whoever executes the steps." This rule explicitly subordinates the skill's default confirmation-seeking behavior to the harness's execution directives.

Rule 5: The Fallback Conflict Resolution

Rule 5—"A rule fights the task" provides the secondary fallback when a skill rule contradicts specific task requirements (lines 126-127 of SKILL.md). This creates a clear hierarchy: harness directives override skill rules, and task requirements override both when explicitly conflicting.

The Architectural Flow of Agent Harness Integration

When a user issues a command through an agent harness, the system processes the request through five distinct control phases:

1. Command Reception: The agent harness receives the request and loads the skill definition from skills/i-have-adhd/SKILL.md.
2. Skill Constraint Evaluation: The i-have-adhd skill applies its formatting rules (lead-with-action, numbered steps, concrete time estimates).
3. Harness Override Application: The harness applies its system prompt directives, such as "always announce tool calls" or "never ask for confirmation," overriding the skill's default "ask before doing" behavior.
4. Response Generation: The final output respects the skill's ADHD-friendly formatting while following the harness's execution directives.
5. Tool Execution: The harness directly invokes tools (file edits, CLI calls) without prompting the user when the system prompt mandates immediate execution.

Configuring Harness Overrides in Practice

Developers configure the precedence relationship through harness-specific configuration files. The repository includes examples for major agent platforms.

OpenAI Harness Configuration

The skills/i-have-adhd/agents/openai.yaml file demonstrates how to force the skill to skip confirmation steps and execute tool calls directly:

# skills/i-have-adhd/agents/openai.yaml

name: i-have-adhd
system_prompt: |
  You are a helpful assistant equipped with the i-have-adhd skill.
  When the skill requests a tool call, announce it and execute it
  without asking the user for confirmation. The skill's output
  formatting rules still apply.
tools:
  - name: edit_file
    description: Edit a file according to a supplied diff
    parameters:
      path: {type: string}
      diff: {type: string}

When the harness processes a skill invocation like:

{
  "role": "assistant",
  "content": "/i-have-adhd edit_file path=README.md diff=\"+Added line\""
}

The system produces output that follows ADHD-friendly formatting while executing immediately:

1. Apply the diff to `README.md` (about 30 seconds).
2. Verify the change with `git diff`.

The harness announces the edit_file tool call and runs it without user confirmation, suppressing the skill's default "ask before doing" behavior per Rule 6.

Gemini and Multi-Harness Support

Equivalent configurations exist for other platforms. The skills/i-have-adhd/agents/gemini.toml file provides the same override structure for Google's Gemini harness, while plugin.json in the repository root registers the skill with various agent types and defines the precedence relationships.

Why Harness Precedence Matters for ADHD-Friendly AI

The precedence model balances safety and autonomy while maintaining consistent output formatting. When the harness overrides confirmation requirements, it enables autonomous execution of safety-critical actions (destructive commands, production deployments) while the skill's ADHD-friendly structure (action-first, numbered steps, concrete time estimates) remains intact. This architecture allows developers to add new agent harnesses by defining system-prompt overrides without modifying the core skill logic in SKILL.md.

Summary

• Rule 6 in SKILL.md (lines 128-129) explicitly states that agent harness system prompts outrank skill rules, requiring immediate tool execution without confirmation.
• Rule 5 (lines 126-127) provides fallback conflict resolution when rules contradict task requirements.
• The skill's ADHD-friendly formatting (lead-with-action, numbered steps, time estimates) persists even when the harness overrides confirmation behavior.
• Configuration files like agents/openai.yaml and agents/gemini.toml define the specific override directives for each harness type.
• The harness controls whether tools execute immediately or await user confirmation, while the skill controls how the response is structured.

Frequently Asked Questions

What happens when an agent harness system prompt contradicts the i-have-adhd skill rules?

According to Rule 6 in SKILL.md, the agent harness system prompt always wins. The skill must announce tool calls when the harness requires them and execute actions immediately instead of asking "want me to," while still applying ADHD-friendly formatting to the response content.

How do I configure the OpenAI agent harness to override i-have-adhd confirmation behavior?

Create or modify skills/i-have-adhd/agents/openai.yaml to include a system_prompt directive that explicitly instructs the agent to execute tool calls without confirmation. The skill's output formatting rules will still apply to the response text, but the harness will control the execution flow.

Does the i-have-adhd skill lose its formatting benefits when overridden by a harness?

No. The precedence rules ensure that while the harness controls whether to execute tools or ask for confirmation, the skill continues to control how the response is formatted. The output still follows ADHD-friendly conventions: leading with the action, using numbered steps, and providing concrete time estimates.

Where are the harness configuration files located in the repository?

Harness-specific configurations reside in skills/i-have-adhd/agents/, including openai.yaml for OpenAI agents and gemini.toml for Gemini integration. The plugin.json file in the repository root registers the skill with various agent harnesses and defines the precedence relationships.

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