Agent Skills for Context Engineering

This repository is an open collection of Agent Skills for context engineering, the discipline of managing what enters a language model's context window. It's aimed at developers building production agent systems who need patterns for context curation, multi-agent coordination, memory design, tool contracts, and evaluation, rather than just prompt wording tricks.

Each skill is a self-contained directory with a SKILL.md file and supporting references, structured so agent platforms load only the skill name and description at startup and pull in full content when a skill becomes relevant to the task. The project is platform-agnostic: skills work across Claude Code, Cursor, Codex, and any host that supports the Agent Skills directory spec or Open Plugins manifest.

Key features

• Foundational skills covering context anatomy, degradation patterns (lost-in-middle, poisoning, distraction, clash), and compression strategies for long-running sessions.
• Architectural skills for multi-agent patterns (orchestrator, peer-to-peer, hierarchical), memory systems (short-term, long-term, graph-based), tool design, filesystem-based context, and hosted/sandboxed agent infrastructure.
• Operational skills for context optimization (compaction, masking, caching), self-managed context editing, harness engineering with rollback and approval boundaries, and self-improvement loops for self-modifying systems.
• Evaluation skills including LLM-as-a-Judge techniques (direct scoring, pairwise comparison, rubric generation, bias mitigation) and deterministic quality-gate design.
• Development methodology skills for scoping LLM projects end to end, from task-model fit analysis to pipeline architecture.
• Worked examples showing how several skills combine in real systems: a personal "digital brain" skill, a multi-agent X-monitoring pipeline, an LLM-as-judge toolkit with tests, and an author-style SFT training pipeline.
• Progressive disclosure design that keeps agent context usage low by loading skill details only when activated.

Ideal use cases

This fits teams building agent systems that need to survive long sessions without losing coherence, or that run multiple coordinated agents and need shared conventions for memory, tool contracts, and evaluation. It's a good fit if you're using Claude Code, Cursor, or Codex and want ready-made skill definitions instead of writing context-management guidance from scratch, or if you're designing an autonomous agent harness and need patterns for rollback, novelty gates, and approval boundaries.

It's also useful as a reference even without installing anything: the skill descriptions and examples document known failure modes (attention scarcity, U-shaped attention curves) and give vocabulary for discussing context design decisions.

It is not a framework, SDK, or runtime. There's no code to import and no agent orchestration engine included. If you want a drop-in multi-agent library or a hosted evaluation service, look elsewhere; this repo gives you the conceptual and structural patterns, and the examples use Python pseudocode rather than installable dependencies.

Installation

Claude Code (plugin marketplace):

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

Then install the plugin:

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

This installs all skills at once; Claude activates them automatically based on task context.

Cursor, Codex, or other Open Plugins hosts:

Clone the repository or add it as a plugin directory. The host reads the manifest at .plugin/plugin.json and discovers skills under the repo-root skills/ directory.

Installing individual skills manually:

Agent Skills require a directory layout, not a flat markdown file. Copy the full skill folder:

# Cursor project-scoped install
mkdir -p .cursor/skills
cp -R skills/context-fundamentals .cursor/skills/

# Claude Code project-scoped install
mkdir -p .claude/skills
cp -R skills/context-fundamentals .claude/skills/

# Codex project-scoped install
mkdir -p .codex/skills
cp -R skills/context-fundamentals .codex/skills/

# Generic repo-scoped install (Open Plugins hosts)
mkdir -p .agents/skills
cp -R skills/context-fundamentals .agents/skills/

Do not flatten SKILL.md into a single file; relative references/ paths will break. For custom agent frameworks, extract the principles directly from each skill's markdown rather than relying on directory discovery.