Give it the ticket, it produces a plan, asks clarifying questions, then implements. Skills encode your repo conventions; hooks gate dangerous commands. Subagents handle reviews and verifications without polluting the main thread.
Claude Code only
If you live in the terminal, you don't need a chat tab next to your editor. Claude Code can plan a change, edit the files, run the tests, and stage the commit — all in one session. CodeRabbit reviews the PR; you bounce findings back to the same Claude Code session for fixes.
Same role as the recommended stack. The CLI-only loop pairs especially well with CodeRabbit because the agent can pipe `gh pr view` straight into its own context to fix findings without you copying anything.
- claude-codeFree tier
- coderabbitFree (OSS)
- claude-code$20 Pro
- coderabbit$12/dev
- claude-code$100 Max
- coderabbit$60 (5 devs)
- 1Open a sessionClaude Code
`cd repo && claude`. Hand it the ticket as the first message. Let it ask the clarifying questions a chat tab would have surfaced.
Prompt · Plan-and-implement from a ticket (Claude Code)I want a small, well-scoped PR for the change below. Plan first, then implement. Ask 1 to 2 clarifying questions before you start writing if anything is genuinely ambiguous. Ticket / problem: """ {{paste ticket}} """ Constraints: - Smallest viable diff. - Match the existing code style. Read 2 to 3 nearby files before introducing a new pattern. - Run the project's test command before reporting done. - If a CLAUDE.md or AGENTS.md exists at the repo root, treat it as authoritative for conventions. When you're done: print the file list, the test result, and a 2-line summary I can paste as the commit message. - 2Open the PRClaude Code
Use Claude Code's `/pr` workflow (or `gh pr create`). It writes the PR description from the diff. CodeRabbit posts its review within ~60s.
- 3Triage CodeRabbit, in-sessionClaude Code
Pipe the review into the same Claude Code session: `gh pr view --json comments | claude`. The agent sees the findings, the diff, and the codebase together — no copy-paste.
Prompt · Triage and patch CodeRabbit findings inside Claude CodeI just opened a PR. CodeRabbit posted comments. I'm pasting them below. For each comment: 1. Decide: real issue, style nit, or wrong (false positive). 2. For real issues + non-trivial nits: produce a minimal patch. Run the tests. Reply on the PR explaining the fix. 3. For false positives: write a 1-line PR reply explaining why we're declining, no code change. Do NOT batch-resolve nits without reading them. Do NOT introduce changes outside the scope of each comment. CodeRabbit comments: """ {{paste 'gh pr view --json comments' or the comment thread}} """ - 4Re-review and mergeCodeRabbit
Push fixes; CodeRabbit re-reviews. Merge once it's quiet and the human review is approved.
Built mostly inside Claude Code, terminal end-to-end. 28+ passes, 40+ PRs, the full SDLC inside one CLI session per ticket.
AGENTS.md at the repo root encodes the conventions Claude Code reads on each session. Open it to see what 'authoritative repo conventions' looks like in practice.
Four loosely-related concerns shipped in one focused PR. Demonstrates the 'plan-and-implement in one CLI session' loop without a separate chat tab.
Constrained, debug-flavored work — the kind of ticket that reads like 'figure out why X is too big and fix it'. Claude Code triaging and patching in the same session.
Ran the same 8 tickets through this stack and the chat-led recommended stack. Time-to-merge dropped from 90 to 55 min/PR average. The trade-off: 2 of the 8 tickets needed a chat-tab debate before the agent could plan well — for those, the chat-led stack still won.
When you genuinely don't know what to build, a CLI agent that's already implementing is the wrong tool. Drop back to a chat tab for the open-ended part, then come back to Claude Code to ship.
Claude Code can run shell commands. Configure permissions and hooks before you hand it a repo with deploy credentials, migrations, or a production .env in scope.
