When to Break the i-have-adhd Rules: 6 Exceptions That Override the Protocol

Break the i-have-adhd rules only in six specific scenarios: explain-style requests, destructive actions, debug spirals, real ambiguity, rule-task conflicts, and harness conflicts.

The i-have-adhd skill enforces a disciplined, action-first communication style designed for users who need clarity without cognitive overhead. Defined in skills/i-have-adhd/SKILL.md, its ten core rules shape every response—numbered steps, concrete next actions, visible wins, and strict elimination of preambles and closers. However, rigid adherence would sometimes work against the user. The source code explicitly carves out six exceptions where breaking the rules serves safety, accuracy, and progress.

Where the Exceptions Are Documented

In SKILL.md (lines 19‑30), a dedicated "When to break the rules" section lists the six overrides. These are not vague guidelines but precise triggers with defined behaviors. The skill preserves its structural identity—still no preambles, still action-oriented—while temporarily relaxing specific constraints.

The Six i-have-adhd Rule Exceptions

1. Explain-Style Requests

When the user asks for a full explanation or walkthrough, the skill expands beyond its usual concise format while maintaining the core shape.

Normal i-have-adhd responses cap lists at five items and suppress elaboration. For explain-style requests, the skill still avoids preambles and closers but provides sufficient detail to teach rather than just direct.

User: "Can you walk me through setting up a virtual environment?"

Assistant:
1. Install Python 3.10 or newer.
2. Run `python -m venv .venv`.
3. Activate the environment (`source .venv/bin/activate` on Unix, `.venv\Scripts\activate` on Windows).
4. Install dependencies with `pip install -r requirements.txt`.
5. Verify the setup by running `python -c "import sys; print(sys.version)"`.

2. Destructive Actions

Operations that could cause data loss require explicit confirmation, overriding the default "act first" imperative.

The skill interrupts its flow to request user approval before executing commands like rm -rf, forced Git pushes, or database schema migrations.

User: "Delete the `tmp` folder."

Assistant:
⚠️ This action will remove all files in `tmp`. Confirm?

(If the user replies "yes", proceed with `rm -rf tmp`.)

This exception prevents irreversible mistakes that would otherwise violate implicit user trust.

3. Debug Spirals

After three consecutive "still broken" turns, the skill stops guessing and asks a diagnostic question.

Without this exception, the i-have-adhd pattern risks an infinite loop of rapid, failed fixes. The rule triggers when the user reports persistent failure despite multiple correction attempts.

User: "My tests still fail after the last change."

Assistant:
We've tried three fixes without success.
What error message do you see when you run `pytest`?

Only one diagnostic question is issued, halting the spiral while maintaining forward momentum.

4. Real Ambiguity

When the request lacks sufficient information, the skill prefers a single clarifying question over guessing.

This exception protects against wasted effort on misinterpreted instructions. The skill identifies genuine uncertainty—distinguishing it from resolvable ambiguity—and seeks specification.

User: "Update the config."

Assistant:
Which configuration file should be updated – `settings.yaml` or `config.json`?

The response contains exactly one question, no speculation, and no preamble.

5. Rule-Task Conflict

If a strict rule would delete the answer itself, the task's content takes precedence.

This meta-exception ensures that rule-following never destroys utility. The response shape (numbered steps, no preambles) persists, but content that would be excised by literal interpretation is preserved when essential to the answer.

6. Harness Conflict

When the surrounding agent harness demands a tool call or modified time-estimate framing, the harness's requirement wins.

The i-have-adhd skill operates within larger systems. If the orchestration layer requires specific output formats—for tool execution APIs or time estimate protocols—the skill adapts while keeping its structural signature intact.

User: "Run the evaluation script."

Assistant:
Running the script now… *(tool call executed by the surrounding harness)*

The [running] indicator preserves the skill's action-first ethos while satisfying external requirements.

How Exceptions Preserve the Core Identity

Each exception follows a consistent pattern: the deviation is minimal, justified, and temporary. The skill never reverts to generic assistant behavior. Preambles remain banned. Closers remain omitted. The numbered, state-rested, win-visible structure endures even when specific rules bend.

According to the SKILL.md source, "These exceptions ensure safety, clarity, and progress when the strict rule set would hinder the user's goal." They are deliberately limited to six scenarios so that the core style stays predictable and the cognitive load of "which voice am I hearing" never increases.

Summary

• Explain-style requests allow expansion for teaching without preambles or closers.
• Destructive actions trigger mandatory confirmation before execution.
• Debug spirals halt after three failed attempts, switching to diagnostic questions.
• Real ambiguity prompts a single clarifying question instead of guessing.
• Rule-task conflicts preserve essential content over literal rule adherence.
• Harness conflicts defer to external system requirements while maintaining structural identity.

Frequently Asked Questions

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

A debug spiral activates when the user reports that a fix failed for three consecutive turns. Rather than attempting a fourth guess, the skill stops and asks one targeted diagnostic question to break the loop. This is defined in SKILL.md lines 19‑30.

Does the i-have-adhd skill ever use preambles or closers during exceptions?

No. Even when breaking other rules, the skill maintains its prohibition on preambles, recaps, and closers. The exceptions modify content depth, safety confirmations, or question-asking behavior—not the fundamental stripped-down structure.

How does the skill handle destructive commands like rm -rf?

The skill interrupts its normal flow to request explicit confirmation, displaying a warning with the specific consequence. Only upon user confirmation does execution proceed. This overrides the default "immediate action" rule to prevent data loss.

Can the surrounding system override i-have-adhd rules?

Yes, through the harness conflict exception. When the agent harness requires tool calls or specific time-estimate formatting, the skill adapts to those requirements while preserving its numbered-step, no-preamble presentation style.

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