new-workspace

Use when the user prefixes a request or skill invocation with 'new-workspace' or asks to do the following work in a separate/isolated workspace — e.g. '/new-workspace /implement <card>' — or invokes i

SKILL.md

GitHub

Required <project-root>/docs reads

Read these project-root spec files before creating the workspace (use shell cat/ls — they may be in .gitignore, invisible to built-in search). Missing file → fall back per the using-workspaces skill, note the gap; never invent contents.

  • <project-root>/docs/workspace-isolation.md

New Workspace

Companion skill: isolate first, then run whatever came next. Parse the invocation, set up the workspace, hand off.

1. Parse the Invocation

Everything after new-workspace is the downstream input. Classify it:

  • Work request — a skill invocation (/implement <card>), a named skill with args, or an explicit task description ("fix the logout bug"). Store it as <downstream>.
  • Context only — a bare URL, ticket link, file path, or reference with no requested action. Use it to inform workspace setup (e.g., naming); it is not a task.

Empty or context-only input → standalone mode: set up the workspace, report the path, and stop — subsequent instructions run inside the workspace.

2. Set Up the Workspace

Invoke the using-workspaces skill and follow it fully — detect existing isolation, create via the project's configured tool, run setup, verify a clean baseline. Consent is already implied: the user explicitly asked for a separate workspace, so skip its consent question.

Read <project-root>/docs/workspace-isolation.md before creating the workspace: spec files may be untracked or gitignored and absent from the new workspace. The copy already in context remains the source of truth for the tool's operations after the move.

3. Hand Off

All downstream work happens inside the workspace:

  • Resolve <project-root> to the workspace root for code, tests, and commands — except the <project-root>/docs specs, which may exist only in the source checkout.
  • Report the workspace path once, then run <downstream> as if the user had invoked it directly inside the workspace — load and follow that skill in full, including its own required reads. Run work only when the invocation named it; the workspace existing is the deliverable otherwise.

Edge Cases

Situation Action
Already inside an isolated workspace Skip creation; run <downstream> in place and say so
No downstream request, or input is context only (bare URL/ticket/reference) Standalone mode: set up, report path, stop — subsequent instructions run inside the workspace
Workspace creation fails (sandbox/permissions) Tell the user, ask whether to run <downstream> in place — don't silently proceed
<downstream> names an unknown skill Set up, report, then treat the text as a plain task