ramonk2o/sensei

A lean maintenance cycle for Claude Code skill plugins. Sensei plans a new plugin, sets up a plugin repository, keeps its backlog and changelog, carries each change to a skill from a real case to a merge, and releases versions. A script handles the checks that are certain, and readers who never saw the conversation handle the ones that need judgment.
The skills are written in English and talk to you in your own language. Everything they write to your repository is in English unless you ask for another language.
## Skills
The names follow a dojo: meditate on a new plugin, prepare the repository, train each skill, log what is left and what was earned, graduate a version.
meditate plans a new plugin from an idea, before any skill exists:
who it serves and the concrete requests it must answer;
which skills it has, where each one's scope ends, and which comes first;
its principles, written to docs/principles.md, and one backlog entry per skill, for prepare and train to pick up.
Run it with /sensei:meditate.
prepare sets up a plugin repository, new or existing:
what's missing of the skeleton, and the principles your skills follow;
a CLAUDE.md with the project's facts;
settings that turn off conflicting plugins and skills, and the installed copy of the plugin itself, in its own repository (except for sensei itself, whose installed copy does the maintenance);
the rules of an old process document moved into place.
Run it with /sensei:prepare.
log keeps BACKLOG.md (ideas, bugs and debt, each from a real case) and CHANGELOG.md (what changed in each version).
train creates or changes a skill:
the real case, rounds of questions with recommended answers, and a change plan you approve;
the edit on a branch, reaching every sibling skill that states the same rule;
then checks, independent reviews, a test scaled to the change, the changelog entry and the merge.
It also judges issues and pull requests, audits consistency and recreates skills from old ones.
graduate releases a version: number, checks, changelog, plugin.json, tag, push, and one zip per skill when your skills stand alone, or one zip of the whole plugin otherwise. Run it with /sensei:graduate.
Two read-only agents review changes for train:
examiner judges a new skill's architecture, or any skill's in an audit that asks for it;
sweeper finds leftovers, contradictions, redundancy, drift between skills, broken references, inconsistent terms, vague instructions and scope overlap.
scripts/check.py checks:
sizes and description lengths against your principles;
names across folders, README and manifests;
the files each skill points to;
the changelog's structure, and that each branch changing the plugin has a changelog entry;
claude plugin validate, on an archive of the last commit.
It also prints a scope map of your skills.
## How a change goes
| Kind | Plan | Branch | Reviews | Test |