expert-instruction_skill

Overview

This skill defines the primary instruction protocol for senior engineering agents, focusing on autonomous reasoning, tiered memory management, and verifiable goal execution. It codifies a cognitive stack and operational norms to convert a large language model into a consistent, testable engineering architect. The goal is reliable, auditable actions with minimal supervision for long-horizon technical work.

How this skill works

The skill inspects project context, dependency state, and active patterns to form hypotheses and simulated outcomes before taking actions. It enforces a five-stage reasoning loop: perception, hypothesis, simulation, action, and criticism, coupled with atomic commits and self-audits. Memory layers (context, working, long-term) are used to persist relevant facts and patterns across sessions and decisions.

When to use it

• When you need an LLM to execute multi-step engineering work with traceable outcomes.
• When managing long-lived tasks that require memory across sessions or components.
• When enforcing coding standards, atomic changes, and verifiable definitions of done.
• When coordinating multi-agent workflows or requiring rigorous context engineering for token efficiency.

Best practices

• Start with a Context Discovery pass to map frameworks, versions, and active patterns before making changes.
• Define a clear Definition of Done (e.g., passing tests, lint clean) and break work into atomic, testable commits.
• Use Search-First approaches (rg/grep) and selective reading with offsets/limits to minimize token noise.
• Always report status and errors; avoid silent failures and never fabricate APIs or specs.
• Maintain working and long-term memory entries for recurring patterns and post-mortem learnings; use /rewind when a path becomes invalid.

Example use cases

• Automating a refactor that spans multiple repositories with consistent style and tests.
• Triaging and fixing a production incident with simulated changes and atomic rollouts.
• Designing an agentic pipeline that coordinates tooling, tests, and deployment with A2A handoffs.
• Encoding canonical ‘gold standard’ examples into prompts to improve downstream agent outputs.