What the repository is
Agent Skills for Context Engineering is a public GitHub repository by Muratcan Koylan, licensed under MIT. It packages knowledge about context engineering, multi-agent architectures and production agent systems as Agent Skills: folders with a SKILL.md file and optional scripts/ and references/ folders. An agent that has the skills installed can apply the guidance when a task calls for it, for example when it designs a memory system or debugs an agent that loses track in long sessions.
Per November 2026 the repository lists 18 skills in six groups:
| Group | Skills |
|---|---|
| Foundational | context-fundamentals, context-degradation, context-compression |
| Architectural | multi-agent-patterns, long-horizon-prompting, memory-systems, tool-design, filesystem-context, hosted-agents |
| Operational | context-optimization, self-managed-context, latent-briefing, evaluation, advanced-evaluation, harness-engineering, self-improvement-loops |
| Development methodology | project-development |
| Cognitive architecture | bdi-mental-states |
Besides the skills, the repository holds six example systems, a “researcher” folder with rubrics and benchmarks for improving the skills, and plugin manifests for several hosts. The README says the work is cited in academic papers, including “Meta Context Engineering via Agentic Skill Evolution”. It also reports router benchmarks of its own; those are self-reported and not independently checked.
How to install it
In Claude Code, as a plugin (all skills at once):
- Add the marketplace:
/plugin marketplace add muratcankoylan/Agent-Skills-for-Context-Engineering - Install the plugin:
/plugin install context-engineering@context-engineering-marketplace. You can also browse to it through the plugin menu and choose Install now.
Plugin skills are namespaced, so they show up as /context-engineering:<skill-name> and do not clash with your own skills.
One skill, by hand:
- Copy the whole folder, for example
cp -R skills/context-fundamentals .claude/skills/. - Put it in
.claude/skills/to share it with everyone who works in the repository, or in~/.claude/skills/for all your own projects.
Other tools: the README describes installing from the Cursor Plugin Directory, copying folders into .cursor/skills/, and for Codex, GitHub Copilot CLI and other hosts that read .plugin/plugin.json, adding the repository as a plugin directory or using .codex/skills/ or .agents/skills/.
The common misunderstanding
Most people expect “the context engineering skill” to make the agent manage its own context better automatically. It does not change how Claude Code compacts or retrieves. A skill is instructions plus reference material that loads when it is relevant. What it changes is what the agent knows about the subject: when you ask it to design a memory layer, split work over sub-agents or shrink a bloated prompt, it can draw on the guidance in the skill. That makes it most useful for people building agents, less for people only using one.
It also costs little to have installed. Claude Code keeps only the skill descriptions in context; the body loads when a skill is used, and its reference files only when the skill points to them. That is progressive disclosure, the same just-in-time principle Anthropic recommends for all context (see What does Anthropic say about context engineering?).
What this means for a team
A skill committed to .claude/skills/ in a repository reaches everyone who works in that repository with Claude Code. That is a real way to share know-how. Its limits are those of files: it is static guidance someone has to update by hand, it stays in that repository and those tools, and it does not learn from what the team corrects. Skills describe how to engineer context; they are not the team’s context itself. For that layer, see How do you set up context engineering for a team? and Context engineering for multi-agent systems.
Where Bombos fits
A skill teaches an agent the method; Bombos holds the material. It is a plugin that gives every AI tool your team uses, Claude Code, Cursor, Codex, ChatGPT, Copilot or your own agents, one shared memory with team knowledge, learned rules and what already sits in mail, documents and repos, with a visible trail of what each answer used. One person can try it first for their own agents, next to the skills they already run.
Related questions
Is the context engineering skill made by Anthropic?
No. It is a community repository by Muratcan Koylan on GitHub. It uses Anthropic's Agent Skills format (a folder with a SKILL.md file), which is why it installs in Claude Code, but Anthropic does not publish or maintain it.
Does installing all the skills fill up Claude's context?
Not much. Claude Code keeps only the skill descriptions in context and loads a skill's full body when it is used. Long reference files cost almost nothing until a skill points Claude to them.
Can I install just one skill?
Yes. Copy the whole skill folder, for example skills/context-fundamentals, into .claude/skills/ in your project or ~/.claude/skills/ for yourself. The repository warns not to flatten SKILL.md into a single file, because that breaks the relative paths to its references folder.
Does it work outside Claude Code?
The repository says the skills are platform-agnostic. It gives install routes for Cursor, Codex, GitHub Copilot CLI and other hosts that read its .plugin/plugin.json, and manual folders such as .cursor/skills/ or .agents/skills/.
