loop schedules recurring autoresearch experiments with /ar:loop, using CronCreate intervals from 10m to monthly. See when to install the loop skill, how to run or stop jobs, and what to verify before relying on it.
This skill scores 70/100, meaning it is acceptable for directory listing but should be presented as a focused, environment-dependent helper rather than a standalone automation package. Directory users get enough evidence to understand when to invoke it and what schedule options it supports, but adoption confidence is limited by the lack of install guidance, support files, and deeper operational safeguards.
- Clear triggerability: frontmatter defines command `/ar:loop` and says to use it when users run `/ar:loop` or ask to run an autoresearch experiment continuously on a schedule.
- Provides concrete usage examples for starting and stopping loops, including experiment names and interval arguments such as `10m`, `1h`, `daily`, `weekly`, and `monthly`.
- Includes a practical interval-selection workflow and maps user-facing choices to cron expressions, reducing guesswork versus a generic scheduling prompt.
- Depends on a surrounding autoresearch setup and CronCreate capability; the repository evidence does not include install instructions or supporting scripts for this skill.
- Operational details are somewhat thin beyond scheduling: there is limited guidance on validation, failure handling, cron cleanup edge cases, or how the recurring experiment actually runs.
Overview of loop skill
What the loop skill does
loop is a scheduling skill for the autoresearch-agent workflow in alirezarezvani/claude-skills. It starts or stops a recurring autonomous experiment loop for a named experiment, using CronCreate to run that experiment on a selected interval. The core job is simple: take an experiment such as engineering/api-speed, choose a cadence, and create a scheduled job so the agent can keep revisiting the experiment without a manual prompt each time.
Best fit for Scheduled Jobs and recurring experiments
The loop skill is best for users who already have autoresearch experiments defined and want repeatable Scheduled Jobs: rapid checks every 10 minutes, hourly background runs, daily overnight experiments, weekly reviews, or monthly slow-cycle research. It is especially useful when the cost of forgetting to rerun an experiment is higher than the cost of letting an agent revisit it on a predictable schedule.
What makes loop different from a normal prompt
A normal prompt can ask an agent to “keep checking this,” but it does not reliably create a persistent schedule. The loop skill gives the agent an explicit command, /ar:loop, a small argument format, fixed interval choices, and a stop command. That structure reduces ambiguity around cadence, experiment identity, and lifecycle management.
Adoption notes before installing
Install loop only if your environment supports the broader Claude skills workflow and has access to scheduling through CronCreate. The repository path is engineering/autoresearch-agent/skills/loop, and the primary file to inspect is SKILL.md. There are no extra scripts, rules, resources, or reference files in this skill directory, so its behavior depends heavily on the surrounding autoresearch-agent conventions and available experiments.
How to Use loop skill
loop install context
A typical directory install command is:
npx skills add alirezarezvani/claude-skills --skill loop
After installation, confirm that the skill is available as /ar:loop and that your agent runtime can create scheduled jobs. The upstream SKILL.md does not include a standalone installer or helper script, so you should treat this as a command skill inside the larger claude-skills repository rather than a separate CLI application.
Basic loop usage commands
Use the command with an experiment path and, optionally, an interval:
/ar:loop engineering/api-speed
/ar:loop engineering/api-speed 10m
/ar:loop engineering/api-speed 1h
/ar:loop engineering/api-speed daily
/ar:loop engineering/api-speed weekly
/ar:loop engineering/api-speed monthly
/ar:loop stop engineering/api-speed
If you omit the experiment, the skill is designed to list experiments and let you choose. If you omit the interval, it should present a menu. The supported interval mapping is intentionally narrow: 10m, 1h, daily, weekly, and monthly.
Turn a rough goal into a strong prompt
A weak prompt is: “Run this experiment regularly.”
A stronger loop usage prompt is:
/ar:loop engineering/api-speed daily
Use the existing engineering/api-speed experiment. Schedule it as a daily background run.
If a loop already exists for this experiment, tell me before creating a duplicate.
Summarize the cron schedule and how I can stop it.
This improves output quality because it names the experiment, chooses a supported cadence, asks for duplicate-loop awareness, and requests operational confirmation. For rapid investigations, use 10m only when you intend to watch the results closely; for unattended monitoring, prefer 1h or daily.
Files to read before relying on it
Start with SKILL.md in engineering/autoresearch-agent/skills/loop. Check the command frontmatter, the usage examples, and the interval table. Then inspect the broader autoresearch-agent structure in the repository if available, because loop assumes experiments already exist and can be resolved. Since the skill directory has no support files, the most important verification is not hidden implementation detail; it is whether your agent runtime actually supports scheduled execution through CronCreate.
loop skill FAQ
Is loop only for autoresearch experiments?
Yes, in practice. The skill is written for the autoresearch-agent pattern and expects an experiment name such as engineering/api-speed. You may adapt the idea elsewhere, but the command itself is not a general-purpose cron editor.
When should I not use loop?
Do not use loop for one-off research, unsafe automation, expensive jobs without budgets, or workflows that require human approval before each run. Also avoid 10m loops for tasks that create noisy commits, API calls, or notifications unless you are actively supervising them.
How is loop for Scheduled Jobs different from cron by hand?
Manual cron gives you full control but requires you to write and manage cron entries yourself. The loop skill is higher-level: it turns a known experiment plus an allowed cadence into a scheduled agent job. You trade flexibility for safer defaults and faster setup.
Is the loop skill beginner-friendly?
It is beginner-friendly only if the surrounding autoresearch setup is already working. The command syntax is simple, but beginners may be blocked by missing experiments, unavailable CronCreate, duplicate schedules, or uncertainty about where results are stored. Read SKILL.md first and test with a low-risk experiment.
How to Improve loop skill
Give loop clearer experiment inputs
The most common failure mode is an ambiguous or missing experiment. Use the exact experiment path, for example engineering/api-speed, and add context if names are similar. If the agent must choose from a list, ask it to show the selected experiment before scheduling.
Choose intervals by operational risk
Cadence affects cost, noise, and usefulness. Use 10m for active observation, 1h for short background monitoring, daily for overnight learning, weekly for longer trend checks, and monthly for slow-moving experiments. A better loop guide prompt explains why the interval fits the experiment instead of choosing a cadence casually.
Ask for confirmation and stop instructions
After creating a loop, request the cron expression, human-readable schedule, experiment name, and stop command. For example:
After scheduling, confirm the experiment, interval, cron expression, and exact command to stop the loop.
This makes the Scheduled Jobs lifecycle visible and reduces the risk of forgotten automation.
Iterate after the first scheduled run
Review the first run before trusting the loop long term. Check whether the experiment produced useful output, whether the cadence was too frequent, and whether failures were reported clearly. If results are noisy, stop the loop with /ar:loop stop <experiment> and restart with a slower interval or a more focused experiment definition.