Self-Improving Agent

Build agents that learn from user corrections by updating and following dated rules to improve performance and reduce repeated mistakes over time.

Security Scan

The skill's purpose (learning from corrections by maintaining RULES.md) is plausible, but the instructions implicitly require reading/writing workspace files and accessing external resources (calendar, projects folder, sending email) without declaring those resources or describing sanitization — that mismatch is concerning.

Purpose & Capability

The name/description match the SKILL.md: it aims to log corrections and evolve a RULES.md operating manual. However, the guidance expects the agent to consult a 'projects/' folder and check a calendar before suggesting times — capabilities that go beyond the stated, self-contained 'rules' purpose and are not declared as required resources.

Instruction Scope

SKILL.md directs the agent to create, read, update, and silently scan RULES.md and to modify AGENTS.md. It also tells the agent to check calendar data and include status from a projects/ folder when referencing projects. Those are file-system and external-service actions (and could capture user corrections that contain sensitive data). The instructions use the word 'silently' for periodic scans, which grants the agent ongoing, background discretion to read/write without explicit user prompts.

Install Mechanism

This is an instruction-only skill with no install spec and no code files — low installation risk. Nothing is downloaded or written by a package installer.

Credentials

The skill declares no required env vars or config paths, yet it implies access to calendars, project folders, and email workflows. That mismatch means the skill expects capabilities (read/write workspace files, access to calendar/email) but doesn't declare them, so it's unclear what credentials or filesystem access will be used.

Persistence & Privilege

While always:false, the instructions explicitly require persistent behavior: log every correction, scan RULES.md every ~10 interactions, and automatically update AGENTS.md. That grants ongoing write/read persistence in the agent environment and could produce long-lived files containing user corrections (which may include sensitive info).

What to consider before installing

Before installing, consider that this skill will create and maintain RULES.md and modify AGENTS.md in the agent's workspace and expects the agent to read project folders and check calendars. Ask where RULES.md will be stored and who can read it; ensure file permissions and retention policies are appropriate. Require sanitization rules (strip PII or secrets) before logging corrections. If the agent will access calendars or send emails, only enable those integrations with explicit credentials and review what data it may read. Prefer running this skill in an isolated/test workspace first, or require explicit user confirmation before any automatic writes or background scans. If you need stronger guarantees, request the author add: explicit config for storage location, an allowlist of folders the skill may access, a sanitization policy for logged corrections, and a clear declaration of any external services or credentials it needs.

Self-Improving Agent Skill

Trigger

Build agents that learn from corrections and get better over time.

Trigger phrases: "self-improving agent", "agent learns", "correction loop", "agent keeps making mistakes", "teach my agent"

The Correction Loop

User corrects agent → Agent logs correction to RULES.md → 
Next session, agent reads RULES.md → Agent avoids the mistake →
Over time, RULES.md becomes a refined operating manual

Implementation

RULES.md Structure

# RULES.md — Self-Improving Operating Rules

## Communication
- [2026-03-15] Never use "I hope this helps" — just end the message
- [2026-03-18] When drafting emails, provide ONLY the email text — no commentary

## Operations  
- [2026-03-16] Check calendar BEFORE suggesting meeting times
- [2026-03-20] When referencing a project, include status from projects/ folder

## People
- [2026-03-17] Client X prefers formal communication
- [2026-03-19] Always CC studio manager on client emails unless told otherwise

Rules for Rules

• Date-stamp every rule
• One rule per line — atomic, independently useful
• Max ~150 rules (beyond this, models start losing adherence)
• Review monthly: remove stale rules, merge duplicates
• If two rules contradict, the newer one wins
• Promote patterns (not incidents) — "always check X before Y" > "that one time X broke"

AGENTS.md Integration

Add to your AGENTS.md:

## Self-Improvement
After ANY correction from the user:
1. Log the correction pattern to RULES.md with date
2. Identify the general rule (not just the specific instance)
3. Check if a similar rule already exists — update rather than duplicate
4. Silently scan RULES.md every ~10 interactions for contradictions

Metrics

• Track correction frequency over time (should decrease)
• Track RULES.md size (should grow, then plateau)
• Track unique vs repeat corrections (repeats should approach zero)