Recipe for a run that survives Fenil being away. Every mechanism here exists in Claude Code 2.1.245; nothing below needs him watching.
Before he leaves
Batch what only he can supply (credentials, account logins, business facts, unbudgeted spend) into ONE
AskUserQuestion; decide everything else yourself.askUserQuestionTimeoutisnever, so a question asked after he leaves blocks until he is back; ask now or decide yourself.State the stopping point explicitly. Default: ship end-to-end (commit, push, PR, merge, deploy) unless he names an earlier stop.
Set the goal so the session knows when it is done:
/goal <done criteria>; PR merged and deployedExample:
/goal every item in the plan implemented, bun run typecheck && bun run test green, PR merged with verification in the body, deploy green.Tasks: create one per plan item with
TaskCreate. The Stop hook refuses to end the turn while tasks are open, which is what his "continue" prompts used to do by hand.
Running it
- Interactive, named so it can be found later:
claude -n <name>then the prompt. - Non-interactive with a hard spend cap:
claude -p "$(cat plan.md)" --max-budget-usd 30 --dangerously-skip-permissions - Periodic check (real interval, this syntax):
/loop 30m check local test status and new PR comments; act on anything actionable - Usage limit:
autoContinueAtUsageLimitis on in settings, so the session resumes when the window resets. - Resume a run from its PR:
claude --from-pr <number>
When he is back
One report, once: before/after (measurable), what changed, what was deliberately left alone, what needs his hands. The PR body carries the verification commands and their results.