Extensions

Walkthrough

PointmanOfficial

Show people how in a shell on their screen, one step at a time, and take over or review theirs.

Install

pointman add pointman-ai/pointman-walkthrough

On every machine of your workspace, or the ones you pick. Or open it in the app:

Open in PointmanView on GitHub

walkthrough

A Pointman behaviour for teaching at a terminal. An agent opens a shell on its person's screen and takes them through something one step at a time while they watch. It can also take over a shell of theirs, or look at what they ran and say what went wrong.

pointman add walkthrough

How it works

A behaviour is guidance, with no process. Installed on a node, SKILL.md's body goes into every agent run there, on Claude Code, Codex and Gemini CLI alike.

  • Showing how ("how do I…", "show me", "walk me through"): the agent opens one shell on the person's machine, in the project's folder, and says in one line what it'll go through. Then one step at a time: a line on what the next command does and why, the command, a wait for the prompt, and a read of what came back. Steps to learn go one command each, never chained with &&. If a step fails, it says so and fixes it in the open. It ends with the commands, in order, in the thread, and leaves the shell open for them.
  • Taking over ("take over", "finish this"): it reads the shell first, far enough back to see what they did, and carries on from there.
  • Reviewing ("have a look", "what went wrong"): it only reads, and never types, even where it may. It answers in the thread: what happened, and what to run next.
  • Never in their shell, without asking first: anything destructive or hard to undo (deleting, force-pushing, dropping data, sudo), secrets typed in plain text, or long jobs that hold their terminal. If they take the keys back, it stops.

What agents use

Pointman's terminal tools: open_pane(shell={machine, folder}) opens a shell on the person's screen, then terminal_wait, terminal_send and terminal_read work in it. A person can let an agent read or type in a shell of their own too, and read_layout shows each shell and what the agent may do in it.

Commands and settings

None. It's guidance only.

Files

File What
extension.toml The manifest: [behavior], for Claude Code, Codex and Gemini CLI
SKILL.md The guidance, in the standard skill format (frontmatter, then the body), so it works as an ordinary skill outside Pointman too

Releasing

Bump version in extension.toml, commit, and make a GitHub release (not a draft or prerelease) tagged v<version>. There's nothing to build, so it has no assets. The marketplace lists it within 30 minutes, and pointman upgrade moves nodes onto it.

Licence

MIT. See LICENSE.

Versions

  1. 0.1.1

    Walkthrough 0.1.1: the first release from its own repo. Same manifest and guidance as before, with a public README and an MIT licence.