pskoett/pskoett-ai-skills

What it solves

This project provides a collection of structured "skills" for AI agents to prevent common failure modes like scope creep, degraded context, and repeating the same mistakes across sessions. It transforms agent behavior from simple prompt-based execution into a reliable, multi-stage workflow with built-in verification and learning loops.

How it works

The project implements two primary feedback loops to improve agent output:

• Inner Loop (Within a Session): Focuses on immediate execution. It uses skills like plan-interview for alignment, intent-framed-agent and context-surfing to monitor drift and context health, verify-gate and self-healing to automatically diagnose and fix failures, and simplify-and-harden for a final quality review.
• Outer Loop (Across Sessions): Focuses on long-term improvement. It uses learning-aggregator to find patterns in session logs, harness-updater to encode those patterns as permanent rules, and eval-creator to turn those rules into regression tests.

These skills can be installed as plugins for various agents (Claude Code, Codex, GitHub Copilot CLI) or as standalone skills following the Agent Skills specification.

Who it’s for

Developers using AI agents for coding tasks who want to move beyond basic chat interactions and implement a professional software engineering workflow (planning, verification, and continuous learning) for their agents.

Highlights

• Self-Healing Capabilities: Automatically diagnoses, patches, and verifies failures mid-task instead of relying on the user to report errors.
• Cros-Session Learning: A system for capturing errors and promoting them to permanent project rules and regression tests.
• Multi-Agent Orchestration: Includes a control-session-orchestrator for coordinating work across multiple agents and sessions.
• Broad Compatibility: Supports installation via GitHub CLI (gh skill), Agent Plugins 1.0, and native plugins for Claude Code and Codex.