create-plugin

name	create-plugin
description	Create minimal, convention-matching plugins or extensions for an existing codebase. Use when the user asks for a plugin, extension, custom module, integration, or add-on for an open-source/project repository and wants the work to stay simple, readable, and aligned with the host project's structure and style.

Create Plugin

Build the smallest plugin that solves the requested problem while fitting the host codebase.

Workflow

1. Learn the codebase before designing anything.

   • Read the GitHub repository first.
   • Read official docs linked from the repository, especially installation, API, examples, and contribution/style guidance.
   • If the deepwiki skill is installed, use DeepWiki MCP for public-repo source-code queries before guessing.
   • Use DeepWiki MCP for three things in particular:
     • structure to map the documentation/wiki layout,
     • contents to read relevant topic pages,
     • ask to query extension points, module organization, APIs, naming patterns, and built-in examples to copy.
   • Ask questions that are concrete and code-oriented, for example:
     • "What are the main extension points for adding a new plugin/module?"
     • "Which built-in modules are the smallest and best examples to study first?"
     • "How is the package organized, and where should a minimal extension live?"

2. Learn the modular structure.

   • Identify the extension points, package layout, naming conventions, imports, base classes, and how existing modules are organized.
   • Use DeepWiki MCP source-code queries to accelerate this discovery when the repo is public.
   • Inspect one or two similar built-in modules or extensions and mirror their shape.

3. Make a short plan before coding.

   • State the minimal implementation approach.
   • Prefer the smallest API and fewest files that still feel native to the project.
   • Call out any deliberate simplifications.

4. Implement the plugin.

   • Match the host project's conventions first: formatting, file layout, typing style, docstrings, imports, naming, and dependency choices.
   • Prefer readability over novelty.
   • Avoid adding dependencies unless clearly justified.
   • Do not modify upstream/vendor code unless the user explicitly asks for an in-tree patch.

5. Check convention fit.

   • Verify that names, module boundaries, and public API feel consistent with the host codebase.
   • If the environment is difficult, it is acceptable to skip full runtime testing, but still do lightweight checks that are feasible, such as syntax checks, lint-style review, or import-path review.
   • Be explicit about what was and was not verified.

6. Document the plugin.

   • Add minimal usage documentation.
   • Explain the core API, state/inputs/outputs if relevant, and the main simplifying assumptions.
   • Keep documentation proportionate to the plugin size.

Design Rules

• Start from existing examples in the host repo before inventing structure.
• Keep the extension minimal and easy to read.
• Prefer a narrow, obvious API.
• Follow the host project's patterns more than generic best practices when the two differ.
• When uncertain, choose the simpler design and note the limitation.
• Do not promise physical/functional fidelity beyond what the implementation actually does.

Output Checklist

Before finishing, confirm all of the following:

• the repository and docs were reviewed first
• the modular structure was inspected
• a plan was made before implementation
• the plugin is minimal and readable
• the code follows host-project conventions as closely as practical
• documentation was added
• verification scope is stated honestly

References

• For a worked example based on CAX, read references/cax-recrystallization-example.md.
• Only read the reference when the current task is similar or when you want an example of how to apply this workflow to a scientific/plugin-style extension.
• If the deepwiki skill is available, read its SKILL and use its ask, structure, and contents commands for public GitHub repositories during the repo-learning phase.