Agent skills
Two files that teach Claude Code or Cursor to write a blockodile module and to connect blockodile to another machine.
blockodile ships skills for coding agents. With them loaded, "write me a module that shows my Docker containers" or "add my VPS and test it" become one-shot requests: the agent already knows the manifest schema, the sandbox rules, the connections file and the two commands that prove an SSH target before you open Settings.
There are two, because an agent picks a skill by its description and one long file covering both jobs got loaded for neither.
| Skill | It covers | Raw file |
|---|---|---|
blockodile-modules | the module folder, the manifest, the web sandbox and the SDK, plus the setup checks and agent detection | /skills/modules.md |
blockodile-connections | connections.json, what a remote machine needs, reading an SSH failure, and proving a target works | /skills/connections.md |
They install themselves
blockodile writes both skills into every agent home it finds on your machine each time it starts,
and keeps them current as the app updates. It only writes where an agent already lives — it will
never create ~/.cursor for someone who does not use Cursor.
| Agent | Paths |
|---|---|
| Claude Code | ~/.claude/skills/blockodile-modules/SKILL.md, ~/.claude/skills/blockodile-connections/SKILL.md |
| Cursor | ~/.cursor/skills/blockodile-modules/SKILL.md, ~/.cursor/skills/blockodile-connections/SKILL.md |
Turn it off in Settings → General → Agent skill, where the same panel has an Install agent
skill button for a manual run. The report names every file it touched. A copy you edited
yourself is kept as SKILL.md.previous before being replaced, never overwritten in place.
Restart the agent
A skill is read at agent startup. Quit and reopen Claude Code or Cursor after the first install.
Install them by hand
Any agent, any machine — each skill is served as plain markdown:
curl -fsSL https://blockodile.dev/skills/modules.md \
--create-dirs -o ~/.claude/skills/blockodile-modules/SKILL.md
curl -fsSL https://blockodile.dev/skills/connections.md \
--create-dirs -o ~/.claude/skills/blockodile-connections/SKILL.mdFor an agent that reads a URL rather than a file, point it straight at those two addresses.
What they cover
- The module folder and the
module.jsonmanifest, both kinds — see Modules. - The web-module sandbox: one file, opaque origin, no relative assets, and the
blockodileAPI. - The traps that cost real debugging time: silently dropped manifests, inlining a library with
String.replace, branching on the OS inside a module. - Connections: the JSON file, the two prerequisites, and one sentence per SSH failure with the command that fixes it.
- Troubleshooting: the setup checks, and why a pane shows no agent status.
They deliberately stop at the blockodile side of the bridge. For driving herdr itself — panes,
tabs, workspaces — herdr ships its own skill: run herdr --skill.