i-have-adhd Rule-Exception Cases: When to Override the Default Formatting Rules

The i-have-adhd agent suspends its standard ADHD-mode formatting in six specific scenarios: detailed explanation requests, destructive actions, debug spirals, ambiguous queries, task-rule collisions, and harness-rule collisions.

The ayghri/i-have-adhd repository provides a specialized AI skill that enforces ten strict formatting rules to generate ADHD-friendly outputs. However, as documented in skills/i-have-adhd/SKILL.md, these defaults are not absolute. The system defines six rule-exception cases where safety, clarity, and execution integrity take precedence over rigid formatting adherence.

The Six Rule-Exception Cases Explained

The canonical definitions for these overrides reside in skills/i-have-adhd/SKILL.md between lines 219 and 236. Each exception triggers automatically when specific runtime conditions are detected.

1. User Requests an Explanation or Walkthrough

When the user explicitly asks for explanations such as "Explain how this works" or "Walk me through the steps", the agent overrides the default brevity rules. According to SKILL.md lines 219-222, the agent must provide a complete exposition with headers for skimmability while still omitting standard preambles and closings.

This exception prioritizes educational clarity over conciseness.

2. Destructive Action Is Pending

Safety overrides formatting when destructive commands are detected. As specified in SKILL.md lines 223-224, the agent must interrupt the workflow to request explicit confirmation before executing operations like rm -rf, force-pushes, schema migrations, or database table drops.

The agent cannot proceed with data-loss operations without user verification, regardless of formatting constraints.

3. Debug Spiral Detected

When the last three consecutive turns all indicate failure (e.g., "still broken"), the agent recognizes a debug spiral. Per SKILL.md lines 225-226, the system stops the iteration loop, surfaces the most likely faulty assumption, and asks a single diagnostic question rather than continuing with speculative fixes.

This prevents wasted computational effort and user frustration during troubleshooting loops.

4. Real Ambiguity in the Request

If the user's intent cannot be reliably inferred, the agent defaults to clarification over guessing. Lines 227-228 of SKILL.md mandate that the agent asks one specific clarifying question instead of delivering potentially misleading answers based on assumptions.

This exception preserves answer accuracy when context is insufficient.

5. Rule Collides With Task Intent

When obeying a formatting rule would delete or obscure the actual answer, the task intent wins. Lines 229-232 specify that for queries like "what are my options", the agent must present the information in the required shape—whether ranked options, short trade-offs, or recommendations first—regardless of conflicting formatting directives.

The utility of the answer supersedes rigid structure in these collisions.

6. Rule Collides With the Harness

When the surrounding system prompt or tool-harness demands behavior that contradicts a formatting rule, the harness takes priority. As noted in SKILL.md lines 233-236, the agent must follow harness requirements (such as announcing tool calls) even if this violates standard ADHD-mode formatting rules, because the harness's contract is higher-priority than the skill's stylistic preferences.

Source Code Implementation

These exceptions are implemented in the skill definition and mapped through the agent configuration files:

• skills/i-have-adhd/SKILL.md – Contains the canonical rule definitions and the six exception cases (lines 219-236)
• AGENTS.md – Maps skill files to runtime adapters and explains the skill loading mechanism
• README.md – Provides high-level architectural overview
• INSTALL.md – References behavioral expectations during setup

The agent evaluates these conditions at runtime, suspending the ten standard formatting rules only when one of the six exception criteria is met.

Practical Exception Handling Examples

The following conversational patterns illustrate how the agent detects and responds to each exception condition:

Example 1: Explanation Request

User: "Can you walk me through how the OAuth flow works?"

AI: 
1. Client redirects user to `/auth` endpoint.
2. Provider returns an authorization code.
3. Server exchanges the code for an access token.
4. Token is stored and used for subsequent API calls.

The agent provides full headers and detailed steps, overriding default brevity while maintaining the prohibition on opening pleasantries.

Example 2: Destructive Action Confirmation

User: "Run `rm -rf /var/www` now."

AI: "Are you sure you want to delete `/var/www` recursively? This action cannot be undone. Please confirm."

Safety protocols trigger before command execution.

Example 3: Debug Spiral Intervention

User: "The test still fails after the last fix."

AI: "We've tried three attempts and the failure persists. I suspect the test environment may be outdated. Do you want to reinstall dependencies?"

The agent stops the fix loop after three failures to reassess strategy.

Example 4: Ambiguity Clarification

User: "Update the config file."

AI: "Which configuration key should be updated, and what value should it have?"

The agent requests specifics rather than guessing the intended configuration changes.

Example 5: Task-Rule Collision

User: "What are my options for deploying this app?"

AI: 1️⃣ Docker container – quick start, portable.
    2️⃣ Serverless (Vercel) – zero-maintenance scaling.
    3️⃣ Traditional VM – full OS control.
    
    Recommendation: Docker container gives the best balance of speed and flexibility.

The agent provides structured options and a recommendation despite any rules that might favor shorter responses.

Example 6: Harness Compliance

User: "Fetch the latest logs."

AI: "⏩ Calling `fetch_logs` tool now... (tool call announced as required by the harness)."

The agent announces the tool call as mandated by the system harness, overriding formatting rules that might prohibit such meta-commentary.

Summary

• Six defined exceptions override the ten default ADHD-mode formatting rules in i-have-adhd
• Safety first: Destructive actions always require confirmation before execution
• Debug spirals halt after three consecutive failures to prevent wasted effort
• Ambiguity triggers single-question clarification instead of guessed answers
• Task and harness collisions resolve in favor of functional requirements over formatting preferences
• All exceptions are documented in skills/i-have-adhd/SKILL.md lines 219-236 and implemented through the agent runtime defined in AGENTS.md

Frequently Asked Questions

What triggers a debug spiral exception in i-have-adhd?

The debug spiral exception activates when the last three conversational turns all indicate persistent failure, typically signaled by phrases like "still broken" or similar repeated error reports. According to SKILL.md lines 225-226, the agent then stops attempting fixes and instead proposes a diagnostic hypothesis with a single clarifying question to break the loop.

Why does the agent override formatting for destructive commands?

Safety takes absolute priority over formatting rules. As implemented in SKILL.md lines 223-224, the agent must request explicit confirmation before executing any command that could cause data loss, such as rm -rf operations, database drops, or force-pushes. This prevents accidental destruction regardless of the user's ADHD-mode formatting preferences.

How does the agent handle conflicts between rules and system harnesses?

When the system harness or tool-calling protocol requires specific behavior that contradicts ADHD-mode formatting rules, the harness wins. Per SKILL.md lines 233-236, the agent follows the harness contract (such as announcing tool invocations) even when this violates standard output formatting, ensuring compatibility with the broader execution environment.

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