Agent Skills for Context Engineering
A comprehensive open collection of skills for context engineering, multi-agent architectures, and production agent systems.
- Stars
- ★ 18k
- Last updated
- 7d ago
- License
- MIT
- Primary language
- Python
- FA score
- 49/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformCodex · Claude CodeOpenAI API (Partial support)
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- Developers diagnosing agent performance degradation in long sessions can apply context-degradation skills to identify lost-in-middle and other failure patterns.
- Not a fit if
- Teams wanting stability rather than fast-moving experimental self-modifying features
- Users without a host that supports Agent Skills, such as Claude Code or Cursor
- Source review
- 49/100 · Major gaps
What does this agent do, and when should you use it?
This repository is a comprehensive set of Agent Skills for building, optimizing, and debugging production-grade AI agent systems with a focus on context engineering. It contains 17 skills spanning foundational concepts (context-fundamentals, context-degradation, context-compression), architectural patterns (multi-agent-patterns, long-horizon-prompting, memory-systems, tool-design, filesystem-context, hosted-agents), and operational optimizations (context-optimization, latent-briefing, evaluation, advanced-evaluation, harness-engineering, self-improvement-loops). It also includes development methodology (project-development) and cognitive architecture (bdi-mental-states). The repo provides a full examples directory with complete implementations like digital-brain-skill, x-to-book-system, llm-as-judge-skills, book-sft-pipeline, interleaved-thinking, and long-horizon-prompt-lab. It follows the Agent Skills directory spec with SKILL.md files, supports progressive disclosure, and doubles as a Claude Code plugin marketplace. It is cited in academic research on context engineering.
The repository delivers actionable guidance as a set of SKILL.md files with accompanying scripts and references that teach specific context engineering practices. It can be installed via the Claude Code plugin marketplace using commands like '/plugin marketplace add muratcankoylan/Agent-Skills-for-Context-Engineering' and '/plugin install context-engineering@context-engineering-marketplace', or by manually copying skill directories into .cursor/skills, .codex/skills, etc. Skills load full content only when activated, enabling efficient context usage. The examples folder contains complete system designs, such as llm-as-judge-skills (TypeScript, 19 tests), book-sft-pipeline (trains models for $2), and digital-brain-skill (personal OS with 6 modules). Additionally, a researcher directory implements a file-based operating system for continuously improving the skills through automated loops, benchmarking, and skill health scoring.
- Developers diagnosing agent performance degradation in long sessions can apply context-degradation skills to identify lost-in-middle and other failure patterns.
- Teams choosing coordination patterns for multi-agent systems can use multi-agent-patterns to compare orchestrator, peer-to-peer, and hierarchical architectures.
- Engineers building long-running autonomous agents can leverage long-horizon-prompting to write task briefs with success predicates, audit gates, and diversity policies.
- Evaluators needing to assess agent output quality can adopt advanced-evaluation to implement LLM-as-judge techniques like pairwise comparison and bias mitigation.
- Developers designing memory persistence for agents can reference memory-systems to choose between short-term, long-term, and graph-based memory architectures.
- Project leads planning an LLM-based project can use project-development to decide suitability and shape pipelines.
How do you install or deploy this agent?
For Claude Code, add the marketplace with '/plugin marketplace add muratcankoylan/Agent-Skills-for-Context-Engineering', then install the plugin via '/plugin install context-engineering@context-engineering-marketplace' or by browsing and installing the 'context-engineering' plugin. For Cursor or Codex, clone the repo and point the host to the plugin root, or copy skill directories into project-local skills directories (e.g., .cursor/skills/).
How do you use this agent?
Once installed, skills activate automatically based on task context. For manual use, copy individual skill directories (e.g., cp -R skills/context-fundamentals .cursor/skills/) and follow the directory spec — never flatten SKILL.md. Browse the examples folder to see how multiple skills combine in real systems, and consult the researcher benchmarks for performance data on the skill router.
What are this agent's strengths and limitations?
- Comprehensive 17-skill coverage from fundamentals to advanced evaluation, with active examples like llm-as-judge (TypeScript, 19 tests) and book-sft-pipeline ($2 cost).
- Platform-agnostic: works with Claude Code, Cursor, Codex, and any Open Plugins-compliant host, with per-platform install instructions.
- Evidence-driven: includes router benchmarks against frontier models and a skill health gate, showing measurable improvements (+23.4pp on context-fundamentals).
- Active self-improvement loop via the researcher directory, ensuring the skills stay current and are backed by academic citations.
- The breadth (17 skills) may overwhelm users needing a single pattern; the learning curve is non-trivial.
- Installation and activation are most seamless in Claude Code; other platforms may require manual copying and lack auto-activation.
- Some examples and benchmarks use frontier models (e.g., GPT-5.5) that may not be accessible to all users, complicating validation of skills.
- The repo evolves rapidly with experimental self-modifying features, which may be unsuitable for teams seeking stability.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Agent Skills for Context Engineering This agent | 49 · Major gaps | Agent plugin / skillFree + model costs | ★ 18k | 7d ago | Python | Codex · Claude Code |
| Finding-Unknowns Skills | 76 · Good | Agent plugin / skillFree | ★ 353 | 21d ago | Python | Claude Code · Claude.ai |
| Keep the Why | 75 · Good | Agent plugin / skillFree | ★ 231 | 1d ago | Python | Codex · Claude Code |
| PowerContext | 55 · Major gaps | CLIFree | ★ 1.3k | today | Python | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows the repository includes a research loop that claims to never invoke paid LLMs or make outbound writes, with stdlib-only HTTP retrieval and a 1.5MB cap and 30-second timeout. However, these guarantees are self-asserted and not verified. Human approval boundaries and rollback mechanisms exist but are not detailed. Dependencies are pinned but no vulnerability scanning evidence is provided. External effects include deployment workflows and launchd agents, but user confirmation is not explicit. Source attribution is partially present through references and contributor info, but publisher identity is unverified. Deductions: lack of independent verification of security claims, insufficient detail on permissions and external effects.
The repository shows structured skill organization and includes validation scripts and tests, indicating internal consistency. However, dependency availability is not fully verified, and failure messages are not detailed. Deductions: insufficient detail on dependencies and failure handling.
The repository provides detailed installation and usage instructions for multiple platforms (Claude Code, Cursor, Codex) and specifies skill activation scenarios. Capability boundaries are somewhat defined through skill descriptions and activation scenarios, but limitations are not explicit. Environment fit is good, but comprehensive troubleshooting is not provided. Deductions: incomplete capability boundaries and limitations.
Information architecture is clear, with skill directories, examples, and a template. Installation notes are detailed, and naming is stable. Examples and FAQ are provided, but known limitations are not explicitly listed. License is MIT, but versioning and changelog are unclear. Maintenance responsibility is indicated through contribution guidelines and contact info, but update strategy is not explicit. Deductions: unclear known limitations, versioning, and maintenance responsibility.
Output usability is high, with detailed skills and examples. Marginal value is evident, providing unique context engineering skills. Cost-benefit is not detailed, but low-cost examples are mentioned. Deductions: insufficient cost-benefit analysis.
Claim traceability is good, with benchmark results and citations provided. Cross-source corroboration is limited, and separation of facts and inferences is unclear. Deductions: limited cross-source verification and fact/inference separation.
- Publisher identity is unverified; treat with caution.
- Security guarantees (e.g., no outbound writes) are self-asserted and not verified.
- Dependencies are pinned but no vulnerability scanning evidence is provided.
- Known limitations and versioning information are unclear.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v2.3.0 v2.3.0 — Measured Router Benchmark + Corpus-Wide Skill Hardening
Release notes excerpt
The first release with measured benchmark results across four frontier models via the Cursor SDK, a corpus-wide hardening pass across all 15 published skills, and the first full-body Stage 3 effectiveness result. The correction in this release is important: we did not stop at the three descriptions that the router benchmark complained about. Every skill body was audited against the same standard, and the machine-readable substrate moved with it: mechanisms, claims, corpus index, activation fixtures, validators, template, docs, PR/release narrative. …View on GitHub ↗ -
v2.0.0 v2.0.0 — Skills Rewrite: Textbook → Toolbox
Release notes excerpt
What's Changed All 13 skills have been comprehensively rewritten based on learnings from Anthropic's ["Lessons from Building Claude Code: How We Use Skills"](https://www.anthropic.com) article. This is the largest single update to the skills collection. The Core Transformation Skills have been rewritten from textbook voice (explains concepts) to hybrid instructional voice (leads with actions, weaves in reasoning): Before: …View on GitHub ↗ -
v1.1.0 v1.1.0 - LLM-as-a-Judge Skills & Advanced Evaluation
Release notes excerpt
🎉 What's New New Skill: Advanced Evaluation A comprehensive skill for mastering LLM-as-a-Judge evaluation techniques. Based on research from [Eugene Yan's LLM-Evaluators](https://eugeneyan.com/writing/llm-evaluators/). Covers: - Direct scoring vs. pairwise comparison selection - Position, length, and verbosity bias mitigation - Metric selection (Cohen's κ, Spearman's ρ, Kendall's τ) - Production evaluation pipeline design - 10 actionable guidelines for reliable evaluation 📁 [skills/advanced-evaluation/](skills/advanced-evaluation/) New Example: LLM-as-Judge Skills …View on GitHub ↗