compare

Claude Skills and gcontext: not a versus

Skills package instructions. gcontext stores state. They are two layers of the same system.

What a Skill is

A Skill is a SKILL.mdfile with instructions and optional scripts. Claude Code loads it on demand when the current task matches the Skill's trigger.

Skills package procedural knowledge: how to run a deploy, how to review code, how to write tests in this repo's style. You can install community Skills, write your own, or combine several for a single workflow.

What Skills do well

  • Procedural knowledge distribution. Write the deploy process once. Every session that triggers the Skill gets the same steps.
  • On-demand loading. Skills stay out of context until the agent needs them. This keeps the base prompt small.
  • Community sharing. Package a workflow as a Skill folder. Others install it and get the same behavior.
  • Composability. One task can trigger multiple Skills without conflict.

Skills solve a real problem: "How do I make the agent do this task the right way every time?" The answer is a file with steps.

The thing a Skill cannot do

Write. A Skill is read-only at run time. The agent reads the instructions, follows them, and produces output. The Skill file itself does not change.

What the agent learns during the task evaporates when the session ends. Skills tell the agent how to act. They cannot remember what it discovered. Every session starts with the same instructions and zero accumulated knowledge.

State is the other half

Skillsgcontext state
Authored byHumanAgent (and human)
Changes at run timeNoYes
ScopePer-taskPer-instance
DistributionSkill filesGit
StorageSKILL.mdFolder over MCP

A Skill that reads and writes gcontext state

The two layers work best together. A Skill defines the procedure. gcontext holds the context and stores the results.

pattern

The Skill stays the same across sessions. The state folder accumulates knowledge. Each run builds on what previous runs discovered.

skill.md
# deploy-review skill
## steps
1. Read modules/deploy/runbook.md for the current deploy process.
2. Run the deploy checklist against the staged changes.
3. Write findings to modules/deploy/reviews/ with today's date.
4. If the review found blockers, list them. Otherwise, approve.

Questions

Are Claude Skills the same as MCP servers?

No. A Skill is a set of instructions that the agent follows during a task. An MCP server is a service that the agent calls with tools. Skills tell the agent how to work. MCP servers give it capabilities (file access, API calls, state storage).

Can a Skill store what the agent learned?

Not on its own. A Skill file is read-only at run time. But a Skill can include steps that tell the agent to write findings to an MCP server like gcontext. The Skill provides the procedure; gcontext provides the storage.

Should I use Skills or gcontext?

Both. They solve different problems. Skills define how the agent should approach a task. gcontext stores what the agent discovers while it works. One gives instructions, the other gives memory.

Related

Claude Code memory
What persists across sessions and what does not
Self-improving agents
The learnings-file pattern that compounds over time
Subagent coordination
Sharing context between Claude Code subagents
What is agent state?
Definition and management patterns
Instructions plus memory

Install gcontext. Your Skills tell the agent how to work. The state folder remembers what it found.

View on GitHubSee the learnings pattern