How to use Claude to run overnight coding tasks
Ralph Loops coordinates Claude and Gemini CLI workers on four old laptops. Git carries task assignments and results, a manager creates fix tasks, and the author checks an overnight summary in the morning.
news.ycombinator.com
Suggested prompt
Suggested prompt written by Agents In Use, not the source author. Replace bracketed inputs and review tool permissions first.
Complete the task in [Markdown task file] within [repository]. Follow its acceptance criteria, record your changes and any failures, and push the result to [branch]. Review completed work and create separate fix tasks for unmet requirements.
Reach people exploring AI agents
See placements →Evidence at a glance
Sourced account- Source poster / publisher
- HappyVaxman
- Original account
- Read the source ↗
- Published
- Source link checked
- Reported result
- The author reports 95% overnight execution success, automatic detection of about 70% of mistakes, and approximately $15 per month in electricity costs for the four-laptop setup.
- Human involvement
- The author writes task specifications and uses a morning review script; the post acknowledges missed mistakes and repeated iterations.
A linked account supports attribution. It does not mean we independently reproduced its result.
Report an issue with this case
Open correction emailOpens your email app with the case link included. Nothing is sent automatically.
How the agent did it
- Write coding tasks in Markdown and commit them to Git.
- Run start-night.sh to begin the overnight worker run.
- Let workers claim tasks, run Claude or Gemini CLI and push results.
The author built Ralph Loops to put four unused laptops to work on coding tasks overnight. The setup coordinates multiple agents through Git, with laptops connected through Tailscale and workers running Claude or Gemini CLI. The author describes it as batch processing rather than real-time pair programming.
The workflow begins with human-written task specifications in Markdown, committed to Git. Before bed, the author runs start-night.sh. Workers claim tasks, execute them and push their results. A manager agent reviews the work and creates additional fix tasks when failures occur. Heartbeat monitoring exposes agent status, and morning-review.sh provides a way to inspect what happened overnight.
The author reports a 95% overnight execution success rate and estimates that the manager catches about 70% of mistakes automatically. Complex features still require multiple task iterations. Reported electricity costs are approximately $15 per month; the setup also requires existing hardware and patience to configure. The code is MIT licensed.
Who it's for: Developers with spare computers and clearly specified coding backlogs who can review results after an unattended run.
Prompt
Suggested prompt (written by Agents In Use, not the author's)
Source: Hacker News · HappyVaxman
Frequently asked questions
How are tasks assigned and reviewed?
Workers claim tasks through Git, run Claude or Gemini CLI and push results. A manager agent reviews the work and automatically creates fix tasks for failures.
What work does the human still handle?
The author still writes task specifications and runs the morning review script. They also report that complex features require multiple iterations and that the manager catches only about 70% of mistakes automatically.
More details
Remaining steps
- Have the manager review results and create fix tasks for failures.
- Run morning-review.sh to inspect the overnight activity.
About this account
Workers execute coding tasks overnight, while a manager agent reviews results and generates retry or fix work.
Platform: Hacker News. Attribution: https://news.ycombinator.com/item?id=46991243. Prompt: Suggested (written by Agents In Use, not the author's). Sourced by the Agents In Use pipeline.
Credit: HappyVaxman on Hacker News
Source published
Added to Agents In Use
Source checked
Tools and topics
Ralph Loops · Claude CLI · Gemini CLI · Git · Tailscale
Overnight coding · Multi-agent · Git · Task automation
Editorial completeness: 100/100. Measures documented fields, not agent performance or verified outcomes.