Agent Skills for Context Engineering: 4K Stars, Powering Cursor & Codex

Why Context Engineering Beats Prompt Engineering

When the context window of a large language model grows (e.g., 200 k or 1 M tokens), two failure modes appear. The “Lost‑in‑the‑Middle” phenomenon causes the model to retain information at the start and end of the context while ignoring details in the middle. “Attention scarcity” means the model’s limited attention budget is consumed by irrelevant tokens, reducing accuracy on the core task.

Context Engineering is the practice of managing that attention budget. Instead of focusing on a single instruction, it orchestrates all inputs—system prompts, tool definitions, retrieved documents, dialogue history, and tool outputs—so the model receives the smallest high‑signal token set.

Repository Structure

The GitHub repository

https://github.com/muratcankoylan/Agent-Skills-for-Context-Engineering

provides a production‑validated set of Agent Skills organized into three layers.

Foundational Skills

Context Fundamentals – explains the physical laws of context and how to feed data efficiently.

Multi‑Agent Patterns – describes orchestrator, peer‑to‑peer, and hierarchical architectures.

Memory Systems – designs for short‑term, long‑term, and graph‑based memory.

Tool Design – guidelines for creating tools that an LLM can actually use.

Operational Skills

Context Optimization – techniques such as compression, masking, and caching to save tokens and improve accuracy.

Evaluation – systematic methods for measuring an agent’s performance.

Advanced Evaluation (LLM‑as‑a‑Judge) – scoring, pairwise comparison, and generation of evaluation criteria, including using AI to grade AI outputs.

Development Methodology

Project Development – end‑to‑end guide covering task‑model matching analysis and pipeline architecture design.

Platform‑Agnostic Usage

Claude Code

Add the marketplace source:

/plugin marketplace add muratcankoylan/Agent-Skills-for-Context-Engineering

Install the context‑engineering suite:

/plugin install context-engineering@context-engineering-marketplace

After installation Claude Code mounts the expert knowledge base and automatically applies context‑management strategies during complex tasks.

Cursor / Codex / IDE

Download the desired skill files from the skills/ directory (e.g., tool-design or project-development).

Copy the content into rule files:

Global rules – paste core principles into the global rule file.

Project‑level rules – create .cursor/rules/ (Cursor 0.45+) or an AGENTS.md file and insert the SKILL.md content.

When the IDE is asked to design a new feature, it follows the project-development workflow: first analyze task‑model fit, then plan architecture, instead of generating code directly.

Case Studies Included in the Repository

Book SFT Pipeline – walkthrough for fine‑tuning an 8 B model to imitate a specific author’s style, with a total cost of $2.

LLM‑as‑Judge Skills – production‑grade TypeScript implementation containing 19 passing tests that demonstrates how to let AI evaluate AI outputs.