AANA Continuous Self-Improvement Skill

AANA Continuous Self-Improvement Skill provides a structured approach for OpenClaw-style agents to enhance their workflows over time while strictly maintaining user-defined goals, constraints, and safety boundaries. The skill is designed to help agents observe task outcomes, identify mistakes, propose low-risk workflow improvements, and verify these changes, all without altering the agent’s authority, permissions, or operational boundaries. It is particularly intended for scenarios where iterative enhancement of agent performance is desired, but any drift from user intent or expansion of agent autonomy must be avoided.

This skill operates as an instruction-only module and does not install packages, run commands, write files, modify agent instructions, persist memory, or interact with external services independently. Its core principle is to improve workflow processes rather than expanding the agent’s capabilities or permissions. The improvement loop involves observing and summarizing user requests and agent outputs, scoring results against explicit constraints and user preferences, diagnosing actionable causes for any shortcomings, and proposing specific, low-risk improvements for immediate or future tasks. The agent is required to gate any proposed change, ensuring that only those strictly within the current task’s scope and risk profile are applied automatically, while all others require explicit user approval.

AANA Continuous Self-Improvement Skill enforces a set of constraints—physical/factual, human impact, constructed/task, and feedback integrity—to ensure improvements are grounded, transparent, and aligned with user expectations. The agent may suggest better checklists, clearer questions, improved verification steps, safer task orders, or minor wording changes, but must not make persistent changes or expand its scope without user consent. The skill also includes guidelines for reporting and handling review payloads, emphasizing privacy and minimal data sharing. If an AANA checker is configured, only redacted summaries are sent, and the agent must avoid including sensitive information in reports.

Delivered as a plugin within the ClawHub and OpenClaw ecosystems, the skill can be installed via CLI and is governed by the MIT-0 license. Its design is tailored for agents and workflows that require continuous, safe, and user-aligned improvement without compromising authority or security.