Xiaobai Self-Improvement

Self-Improvement Engine - AI Agent that learns from mistakes and continuously improves. No more repeating the same errors.

Security Scan

The skill's instructions to log mistakes and auto-generate/enforce persistent rules fit its stated purpose, but they grant broad, persistent self-modifying behavior with vague boundaries and interactions with other skills—this is coherent but potentially risky and needs clearer safeguards.

Purpose & Capability

Name/description (self-improvement agent) align with the instructions: the SKILL.md describes logging mistakes, generating WORK_RULES.md, and changing behavior based on rules. No unexpected binaries, env vars, or installs are requested.

Instruction Scope

Instructions explicitly read/write persistent files (memory/mistakes.json, WORK_RULES.md) and direct the agent to 'adjust behavior' and enforce rules across sessions. They also call verification commands like 'openclaw cron list' and recommend using 'session_status' or Python without limiting when or how to run external commands. The guidance is high-level and gives the agent wide discretion (e.g., 'Check: Am I about to violate any learned rule? If yes → adjust behavior'), which can lead to unintended or hard-to-audit actions.

Install Mechanism

Instruction-only skill with no install spec or downloaded code — lowest installation risk. Nothing is written to disk by an installer beyond what the agent itself may create at runtime per the instructions.

Credentials

No environment variables, credentials, or config paths are requested. The skill's needs (local JSON/MD files for persistent rules) are proportionate to its stated purpose. That said, the skill will persist data which could include sensitive content if not constrained.

Persistence & Privilege

Although always:false, the skill instructs persistent, cross-session state (mistakes.json and WORK_RULES.md) and self-modification of agent behavior. This persistent, self-enforcing rule set can affect how the agent interacts with other skills and users; the SKILL.md lacks explicit safety controls (consent, limits, sanitization, override/rollback), increasing risk of unintended long-term behavior changes.

What to consider before installing

This skill is coherent with its stated self-improvement goal, but it writes and enforces persistent rules and gives the agent broad discretion to change behavior and call verification tools. Before installing: 1) Confirm where memory/mistakes.json and WORK_RULES.md will be stored and who/what can read them. 2) Require explicit user consent before the skill makes permanent rule changes or enforces rules that affect other skills. 3) Ask for limits: what kinds of data may not be logged (PII, secrets), and provide automatic sanitization. 4) Ensure there's a clear restore/reset/undo mechanism (ability to review and remove rules, and to disable persistence). 5) Prefer more specific, auditable verification steps (exact commands, when they're run) and a safe fail-open/fail-closed policy. 6) If you plan to allow autonomous invocation, consider restricting it until you can audit WORK_RULES.md contents and confirm no sensitive data is being logged. If these safeguards are not added, treat the skill as high-risk for unintended persistent behavior.

Self-Improvement Engine 🔄

AI Agent that learns from mistakes and continuously improves. No more repeating the same errors.

The Problem

AI Agents make the same mistakes repeatedly:

• Forgot to verify before saying "done"
• Made incorrect time calculations
• Promised tasks but didn't execute
• Gave vague responses instead of concrete action

Each session starts fresh, learning nothing from past failures.

The Solution

Mistake Log

Track every mistake in memory/mistakes.json:

{
  "mistakes": [
    {
      "date": "2026-04-21",
      "type": "fake_completion",
      "description": "Said 'cron task created' but only wrote to file",
      "impact": "User didn't receive scheduled messages",
      "lesson": "Always verify task exists via cron list before claiming done",
      "rule": "MUST run verification command after any creation action"
    },
    {
      "date": "2026-04-21",
      "type": "time_calculation",
      "description": "Said '1 minute later' when it was actually 1 hour",
      "impact": "User got wrong expectation",
      "lesson": "Never calculate time mentally - always use tools",
      "rule": "Use session_status or date command for time, never mental math"
    }
  ]
}

Improvement Rules

Based on mistakes, auto-generate rules in WORK_RULES.md:

## Learned Rules (from mistakes)

1. After creating any cron task → MUST run `openclaw cron list` to verify
2. Time calculations → MUST use `session_status` or Python, never mental
3. Multi-step task → MUST checkpoint after each step
4. Tool error → MUST report immediately, don't silently retry

Self-Audit Protocol

Every heartbeat or session start:

1. Read mistakes.json
2. Read WORK_RULES.md
3. Check: Am I about to violate any learned rule?
4. If yes → adjust behavior
5. If I make a mistake → log it immediately

Mistake Categories

Type	Description	Prevention
fake_completion	Said done but didn't do	EVR: Execute-Verify-Report
time_calculation	Wrong time math	Use tools, never mental
silent_failure	Tool failed, didn't tell user	Report immediately
vague_promise	"I'll do it" without action	Execute first, then report
forget_context	Lost important info	Use memory tools
wrong_assumption	Guessed instead of checking	Verify before claiming

Usage Examples

When you make a mistake:

User: "You said 1 minute but it's actually 1 hour"

I acknowledge this mistake. Let me log it:
- Type: time_calculation
- Lesson: Always use tools for time, never mental math
- New rule added to WORK_RULES.md

I will not make this mistake again.

When starting a task:

I need to create a cron task.
Checking WORK_RULES.md:
- Rule #1: After creating cron task → MUST verify with cron list

[Creates task]
[Runs openclaw cron list]
[Confirms task exists]
Task verified and confirmed. ✅

When user corrects you:

User: "You should have checked before saying it was done"

Logging mistake:
- Type: fake_completion
- Lesson: Verify before claiming done
- This becomes a permanent rule in WORK_RULES.md

This mistake is now in my permanent memory and won't happen again.

Integration with Other Skills

• EVR - Self-improve reinforces Execute-Verify-Report
• Memory Guard - Mistakes persist across sessions
• Daily Reflection - Review mistakes at day end
• Error Recovery - Log errors that led to mistakes

Key Principle

Every mistake is a learning opportunity. Log it, learn from it, never repeat it.

License

MIT