Claude /loop: my use cases

Claude Code's /loop runs a prompt or slash command on a recurring interval. Once you have slash commands for common ops tasks, it composes into something useful.

GPU Cluster Monitoring

/loop every 5m and /check-cluster-health and /check-training-status and DM me on slack for errors

• Runs every 5 min, calls health + training status commands, Slack DM only on errors
• No noise when everything's fine
• Slash commands are the key: /check-cluster-health knows to check node status, GPU util, job queues

Periodic Training Status Reports

/loop every 1h to /check-training-status and report on slack #training-status channel

• Regular status updates for team visibility
• Channel becomes a timeline of the run

Production Release Monitoring

get a baseline from /monitor-prod-services, then make the release, after that /loop every 3m to /monitor-prod-services and report if any new issues pop up from the release

• Capture pre-release state (error rates, latency, existing alerts)
• Loop diffs each report against baseline, only flags new issues
• Filters out pre-existing noise

Overnight Experiment Runner

plan experiments for me, then /loop every 1h to invoke the mlx-bench command, analyze each result and decide what to test next based on Darwinian selection

• Claude plans initial candidate set, runs experiments, reads results, decides what to keep/discard, queues next run
• Small wrapper around mlx-bench reads from experiment list, picks up next one each invocation
• Wake up to completed sweep with analysis

What Makes This Work

1. Slash commands for common tasks — /loop is just a scheduler. Value comes from composing it with commands that know your systems.
2. Structured output — Claude needs to parse results and make decisions. Machine-readable summaries > raw log dumps.
3. Clear continuation condition — for experiment loops, need a way to pick up state between invocations (e.g., mlx-bench wrapper maintains a queue file).

Related Work

autoresearch (Karpathy)

• Most direct parallel to overnight experiment runner
• Give agent a train.py + fixed 5-min budget per run, let it modify/train/measure/keep-or-discard, repeat ~100 experiments
• Single session — context accumulates across experiments
• Key difference: fixed time budget makes every experiment comparable. My /loop approach is less structured — Claude decides via Darwinian selection. autoresearch is tighter by design.

The ralph loop (Geoffrey Huntley)

• Broader: monolithic autonomous process for full software dev cycle (allocate specs, execute, verify, repeat)
• Parallel to my ops monitoring loops — continuous observe/evaluate/act without human intervention
• Difference: scope and trust. I use /loop for bounded tasks where I've scripted what "good" looks like; ralph loop owns the entire lifecycle.

Structural difference: /loop and autoresearch run in a single session (context accumulates). Ralph loop starts fresh each iteration (relies on spec for context). Fresh = more robust (no context exhaustion). Accumulated = agent notices trends across iterations.