Decision guide· Learn
Skill vs Prompt, Rule, MCP, Plugin, or Subagent
A decision guide for choosing the right agent abstraction for repeatable AI workflows.
intermediate7 min readUpdated May 4, 2026
Short answer
Use a prompt for one-off work, a rule for broad repo behavior, a skill for a repeatable workflow, an MCP server for structured external access, a plugin for distribution, and a subagent for delegated parallel work.
Quick comparison
| Use this | When it fits | Avoid when |
|---|---|---|
| Prompt | The instruction is one-off or exploratory. | You repeat the same workflow often. |
| Rule / AGENTS.md | The guidance should apply across a repo or team. | The workflow is narrow and task-specific. |
| Skill | A repeatable workflow has a clear trigger and output. | Live external integration is the core value. |
| MCP server | The agent needs structured access to systems or data. | Static instructions are enough. |
| Plugin | You want to distribute reusable skills or apps. | The skill is local or repo-specific. |
| Subagent | Work can be delegated or parallelized. | A simple linear workflow is enough. |
When a skill is the right fit
Skills shine when the user can describe a recurring job and the agent can follow a consistent sequence. Examples include review passes, QA loops, release communication, documentation updates, and investigation protocols.
- The trigger is easy to describe.
- The output shape is stable.
- The workflow has known safety boundaries.
- The agent benefits from task-specific examples or checklists.
Skill authoring is not the same as distribution
A skill folder is the authoring format for the workflow. A plugin is a distribution package that can bundle skills, app mappings, MCP configuration, and presentation assets.