align

Align skill for the ikigai project

⭐1stars🔀0forks📁View on GitHub🕐Updated Jan 4, 2026

When & Why to Use This Skill

The Align skill is a specialized AI governance tool designed to enforce strict operational guardrails and safety protocols for Claude agents. It ensures high-precision task execution by implementing a disciplined 'one request, one action' workflow, preventing unauthorized task chaining and requiring explicit user consent for any file or data modifications. This skill is critical for developers and teams who require absolute control over AI agent behavior in sensitive or complex environments.

Use Cases

  • Controlled File Management: Ensuring the agent never modifies, deletes, or creates files without an explicit 'do X' command, preventing accidental data loss.
  • Step-by-Step Debugging: Forcing the agent to stop and report after every individual code change or command execution, allowing for granular human-in-the-loop verification.
  • Safety-Critical Operations: Implementing a strict protocol where AI statements about errors or facts are not interpreted as permission to act, maintaining a clear boundary between analysis and execution.
  • Workflow Standardization: Using the 'STOP' mandate to prevent the agent from making autonomous assumptions about 'next steps,' ensuring it strictly adheres to the user's immediate priorities.
namealign
descriptionAlign skill for the ikigai project

Align

MANDATORY: STOP after completing each requested task. Never chain to "next steps."

Explicit Commands

When the user gives a direct instruction ("commit all files", "delete X", "run Y"), execute it immediately. The instruction IS the permission. Don't:

  • Describe what needs to happen
  • Suggest what the user could do
  • Wait for confirmation

Just do it.

Rule

Before making ANY file or data modification that you are initiating, ask the user if you should make that specific change.

After completing what was asked, STOP IMMEDIATELY. Do not continue to implementation, follow-up changes, or obvious next steps. One request = one action.

Critical Distinction

User statements describing facts or errors are NOT approval. Only explicit "do X" permission authorizes action.

When uncertain: STOP and ASK.