PullPlane gives Claude the same card brief your team is editing, runs the task away from local machines and production, and brings the result back as work your team can inspect before merge.
How PullPlane runs Claude
Standard Claude coding runs use the Claude Agent SDK. The run receives the card brief and connected repository inside its own isolated sandbox, where it can inspect the codebase, implement the task, and run relevant checks.
The ultracode effort path is the only Claude path that attempts to invoke the Claude Code CLI. If that path is unavailable or fails, PullPlane falls back to the SDK at a high reasoning setting. Standard Claude runs should not be described as Claude Code CLI runs.
Good tasks for Claude
Claude is a fit for implementation work with a clear outcome, constraints, acceptance criteria, and verification plan. PullPlane's UX discovery starter recipe also assigns the implementation step to Claude after research has documented the interaction and accessibility risks.
- Build against a live, team-authored card spec
- Inspect terminal output, touched files, the full diff, and a running preview
- Continue on the card branch when a later run needs to refine the work
The human decision stays at merge
A completed coding run can create a branch and pull request, but it does not merge itself into the default branch. The team reviews the implementation and any required Reviewer, QA, or Security checks before approving the merge.