Build a plugin with your coding agent
Paste one brief into Claude Code, Cursor, Copilot or another coding agent, and it builds, validates and documents a plugin for you.
A coding agent can build a plugin for you. The example plugins repository holds the rules agents follow, in AGENTS.md: every manifest field and its allowed values, the triggers, tools and outputs CoderHelm supports, safe defaults, and a checklist the agent must pass. Most coding agents read AGENTS.md on their own; the brief below tells them to anyway.
Get a copy of the repository
git clone https://github.com/CoderHelm/coderhelm-plugins my-plugins
cd my-plugins
npm ciOr click Use this template on GitHub to get your own copy, public or private.
Paste the brief
Open your coding agent in that folder, fill in the first four lines, and paste:
Build a CoderHelm plugin in this repository.
What it should do: <one or two sentences>
When it runs: <a person's brief | a schedule, such as "Mondays 9:00 UTC" | each new release tag>
What it may touch: <the services it reads, and anything it may change, such as "reads GitHub; may post to Slack">
Where results go: <drafts to review in CoderHelm | a channel, issue or service, after approval>
How to work:
1. Read AGENTS.md in full first. It is the source of truth for every field,
trigger, tool and limit. Use nothing it doesn't list.
2. Pick the example under plugins/ closest to this request and copy it to
plugins/<new-name>/ as your starting point. Say which one you picked.
3. If the request needs something CoderHelm doesn't support (for example a
trigger other than manual, schedule or github.release), choose the
closest supported way, and say what you chose and why in the plugin's
README and in your report.
4. Ask me only what you can't answer from the repository or from the four
lines above. Put anything a team would type (channels, repositories,
accounts, style guides) in settings instead of asking me.
5. Start read only. Make every tool that changes something outside
CoderHelm consequential, so it waits for a person's approval.
6. Write plugin.yaml, the agents, any skills, and README.md as AGENTS.md
describes.
7. Run `node tools/validate-plugin.mjs plugins/<new-name>` and
`npm run check`. Fix and run again until both pass. Don't change tools/,
scripts/ or other plugins.
8. Finish with a short report: which example you started from, what the
plugin does, its tools with their risks, the validator's last line, and
what I must connect, fill in and approve in CoderHelm to use it.Check what it built
The agent starts from the closest example, asks only what it can't answer from the repository, runs the validator until it passes, writes the plugin's README, and reports what you must connect, fill in and approve. Read plugin.yaml and check that every tool marked consequential is one you want to approve each time.
Install it
Push your repository, add it as a source on the Plugins page, install the plugin, and start it once with a brief on a test team. See Public and private repos.
Two filled-in examples
BUILD_WITH_AN_AGENT.md has the brief filled in for:
- A weekly summary of merged pull requests, posted to Slack. A schedule and a brief trigger, read-only GitHub tools, and a Slack post that waits for approval.
- A changelog entry when a pull request is labeled. CoderHelm has no label trigger, so the agent picks a supported one (a schedule that looks for newly labeled pull requests, plus a brief) and says so in the README.
Both were built from these briefs by a coding agent with no other help, and passed the validator on the first run.