Use coding-agent skills
Use the SKILL.md workflows in skills/ to learn a project, create projects, build, install and remove modules, add pages, configure providers and deploy.
The skills/ directory contains workflows for a coding assistant working on this repository or a derived project. These are not the published agent-skills/ packages used by the optional /app/agent runtime. For that product, see Agent starter and Agent runtime.
Before you begin
- A coding agent that reads
SKILL.mdfiles, such as Claude Code. - The repository checked out, or a project you want to install the skills into.
| Skill | Use it when |
|---|---|
project-context | Starting on a project or an area you have not read: what is there and where it lives |
quick-start | Creating a project with the CLI and reaching a working admin sign-in |
new-module | Adding a module end to end, including its region markers and manifest |
new-page | Adding one admin or workspace page with copy, guards and navigation |
new-agent-plugin | Adding an agent capability as a plugin of the one workbench: Skill, cards, panels and operations |
configure-providers | Wiring real providers through /admin/config/* |
deploy | Deploying to Cloudflare, Fly.io, Railway or a Docker host: picking or switching the platform, storing the generated secrets, migrating and checking /setup |
add-module | Installing or upgrading a module through the CLI's package workflow, with a preview first |
remove-module | Uninstalling a module through the CLI's package workflow: previewing the plan, keeping edits, deciding about data and retiring its Skills |
quick-start, add-module and remove-module drive the CLI's package workflow, which a project cloned from GitHub does not use. In a clone, follow the quick start instead of quick-start; you do not need the other two, because every module is already in the checkout.
Every skill has the same four sections: Inputs, Invariants, Workflow and Verification. It names only commands, paths and routes that exist in the repository.
Install the skills
Nothing to install. The repository commits a link at .claude/skills/<name> for each skill in skills/, so Claude Code loads them when you open the clone.
On Windows, git checks the links out as plain text files unless symbolic links are enabled (git config core.symlinks true before cloning, with Developer Mode on). Without them, use the other tab.
Run a skill
Name the task
Ask the agent for the task in plain words, or name the skill, for example "use the new-page skill to add an admin page for exports".
Provide the inputs
Each skill starts with an Inputs section, such as a module key or the deployment platform. The agent asks for anything missing.
Let it follow the workflow
The agent works through the Workflow steps while keeping the Invariants, such as never inventing ids or reading process.env from feature code.
Check the result
The agent finishes with the skill's Verification section: the commands and pages that confirm the task is done.
Let an agent read the project
Documentation describes Kernlift in general; an agent working on your project needs facts about that project. Two sources give it them.
pnpm kernlift context prints a description of the checkout, read from the source each time it runs, so it stays current as you add routes, plugins and settings. It needs no token, network or database, and writes nothing.
pnpm kernlift context # overview and the topics below
pnpm kernlift context routes # every URL and the file that serves it
pnpm kernlift context validators --json| Topic | What it lists |
|---|---|
rules | The sections of AGENTS.md, summarised |
routes | URLs by area (public, auth, app, admin, api) with their files |
plugins | Plugins in src/plugins, whether this build installs them, their routes and nav entries |
config | Each /admin/config/* page with the server functions and repositories behind it |
env | Variables in .env.example with their level and description |
scripts | pnpm scripts other than validators |
validators | pnpm validate:* checks, which ones validate:all runs, and the known failures |
skills | The skills in skills/ |
all prints every topic at once. The project-context skill tells the agent to start here.
The documentation as Markdown. /llms.txt lists every page of this site with a link to its Markdown, /llms-full.txt is the whole site in one file, and any page is available as Markdown by adding .md to its address, for example /docs/quick-start.md. The Chinese versions are under /zh. Point an agent at https://kernlift.dev/llms.txt when it needs a concept or procedure that is not in the code.
Verify
- In Claude Code, the skills from
skills/appear in the list of available skills. pnpm kernlift contextprints an overview whose plugin list matchessrc/plugins/installed.ts.- In the repository,
pnpm validate:skillsconfirms every command, path and route a skill names exists.
Troubleshooting
The agent does not see the skills
Open the agent in the repository root so it finds .claude/skills. If .claude/skills/<name> is a small text file instead of a link, the clone was made without symbolic links; install with npx skills add kernlift/kernlift.
pnpm validate:skills fails after an edit
A skill names a command, path or route that no longer exists. Update the skill to match the code.
Reference
- Skills: what each skill does, asks for and checks
- Agent starter: the separate runtime product and its published skill packages
- Agent Skills specification: the
SKILL.mdformat
Next steps
Validators and the release gate
Why every capability carries an executable check that asserts behaviour and wiring, and how the release gate decides which checks must pass.
Use Postgres and Supabase
Point the app at any Postgres database, including Supabase and Neon, and build its schema from the migration chain.