Choosing the Right Mechanism for Four Scenarios
For each scenario below, name the single best-fit mechanism from this
subject (/loop, /goal, a desktop scheduled task, a cloud routine, or
Remote Control) and justify the choice against the others that might
seem plausible:
- You're actively reviewing a PR in your terminal right now and want Claude to keep re-checking CI status and addressing new review comments for the next hour or so, without you retyping a prompt every few minutes.
- Every weekday morning, regardless of whether your laptop is even open, you want a summary of overnight production error-rate anomalies posted to Slack.
- You kicked off a large refactor at your desk before heading out, and from your phone on the train you want to check progress and, if needed, redirect Claude without losing anything it's already done.
- You're migrating 40 files to a new API and want Claude to keep working, turn after turn, until every call site compiles and the full test suite passes — without you having to confirm "keep going" after each file.
1. /loop, self-paced (no interval given). This is squarely
session-scoped work: you're already at the terminal, mid-session, and
the right check-in cadence depends on what's actually happening (CI
running vs. quiet). A desktop scheduled task or cloud routine would
start a new session disconnected from the PR review you're already
doing, and /goal is the wrong tool because there's no single
verifiable end condition here — "handle new comments as they arrive"
is open-ended, not a completion state a fast model can judge met or
not.
2. A cloud routine. The requirement "regardless of whether your
laptop is even open" rules out /loop and desktop scheduled tasks
immediately — both require your machine on, and a desktop task
additionally requires the Desktop app running. Only a cloud routine
runs on Anthropic-managed infrastructure independent of your machine's
state. A Scheduled trigger for early morning, with a Slack connector
attached, matches the requirement directly.
3. Remote Control. The key phrase is "kicked off... at your desk" and "without losing anything it's already done" — this is a session that already exists and needs to be reached, not new work that needs to be started. A cloud routine would run a fresh, disconnected session with no access to the refactor already in progress. Remote Control connects claude.ai/code or the mobile app to the exact local session that's already running, keeping execution and filesystem access on the original machine while letting you view progress and send messages from the train.
4. /goal. This has exactly the shape /goal is built for: a
single measurable end state ("every call site compiles and the full
test suite passes") that Claude's own output can demonstrate turn by
turn, with no fixed number of turns known in advance. /loop would be
the wrong fit because it's driven by a time interval, not by whether
the condition holds — you'd either poll too often before enough
progress has been made, or too rarely and let a wrong turn compound.
Note that /goal alone doesn't grant unattended execution — pairing it
with auto mode is what lets each turn's tool calls proceed without a
manual approval prompt per edit.
Share this question