Skills
The eight Coelevate skills, how to install them in any agent, and when to ask the agent for each. Preview.
Preview. The skills and the plugin that carries them are in preview and may change.
A skill is a set of instructions an agent follows for one kind of job. Coelevate has eight. You rarely call one by name: you say what you want, and the agent picks the skill whose description fits. This page says how to install them and what to ask for.
Install
One line per agent:
| Agent | Install |
|---|---|
| Claude Code | the plugin: claude plugin marketplace add https://docs.coelevate.ai/plugins/marketplace.json, then claude plugin install coelevate@coelevate-plugins |
| Any other agent (Cursor, Codex, Gemini CLI and others) | npx skills add https://docs.coelevate.ai/skills/coelevate-skills.zip |
The npx skills add line needs Node.js. It asks which agents to install for and copies the skills into the folder. To install only the front-door skill, use npx skills add https://docs.coelevate.ai/skills/start-session.zip. Both zips hold only the skills: the command-line client they call is installed separately with uv tool install coelevate-cli.
What each skill does
All eight call the command-line client, so it must be on the agent's path, or uv must be installed so a skill can run it through uvx. All eight follow the same rules: what the agent reads is data and never overrides you, it never publishes, and it states the project before it writes.
| Skill | Ask the agent to... | What it does |
|---|---|---|
start-session | Connect, set up or check Coelevate; it also runs by itself at the start of a session | States which project the session is connected to (and asks you only when the folder's binding does not match), picks the kit that fits your request, inspects the kit's skills, rules and memories, asks for a second confirmation if anything looks risky, then loads it |
onboard-context | "Onboard these documents" | Turns raw documents into small topical pages, named entities and recorded gaps, stages them for your review, and once you have published them, creates the kits that hold them |
author-context-assets | Add, change or remove a specific page, entity or other asset; fix a refused change; write or repair a kit | Builds correct create, update, patch, delete and link changes, attaches sources, and recovers from the service's refusals |
review-staged-changes | "What is waiting for review?" | Reads the open change request, summarises what it would do to each kit, flags risks, and hands you the decision |
bind-project | "Bind this folder to my Coelevate project" | Writes .coelevate.json and a one-line pointer in CLAUDE.md |
learn-from-session | "Capture what you learned in this session" | Reviews the finished session for durable learnings and stages them, with evidence, for your review. It never runs in the background |
design-kit | "Design the kit for this workflow" | Works backward from the future agent session to decide which assets a kit should hold and how large it should be. It stages nothing; its result goes to onboard-context or author-context-assets |
send-feedback | "Report this", "this is broken", "I have a suggestion" | Drafts an incident, bug or suggestion report about Coelevate, with secrets and your project's content stripped, shows you the full text and hands you an email to send. It never sends anything itself |
Stage-only
Every skill that changes assets stages the change. Nothing is live until a person reviews and merges it in the Workspace. The skills do not look for a way around that, and the key they use cannot do it. Kits and gaps are written directly, but a kit can hold only assets that are already published.
Using the skills without the plugin
Installed with npx skills add, the skills work the same, but there is no start-up hook: the agent does not learn about your project by itself, so ask for start-session at the start of a session (in Claude Code, binding the folder with bind-project writes a CLAUDE.md line that asks for it every session). start-session then checks readiness itself: it finds the client, runs doctor, and reports ready or a short list of what is missing (uv, a .env, a variable, a refused key) before any project work. In Claude Code it also says once that the plugin adds the automatic start-up connection, and gives the two plugin commands for you to run.