AIToday
Large Language ModelsAI Coding AssistantsZenn AI/MLPublished: Oct 11, 2026, 22:00 JST

Weekend dev builds Claude Code skills to beat CodeRabbit limits

Weekend dev builds Claude Code skills to beat CodeRabbit limits

A developer working solo on weekends built Claude Code skills, including /coderabbit-queue, which lines up pull requests and re-requests CodeRabbit reviews about every 70 minutes, since his Essential plan sometimes allowed only one review per hour.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Context & Analysis

The author's context is a split working life: he uses AI tools largely without worrying about volume at his day job, but pays for his own hobby development, holding the Pro plan for Claude (including Claude Code) and the Essential plan for CodeRabbit. That asymmetry is what shaped the design. When weekend batch sessions kept hitting CodeRabbit's review cap, he turned the constraint into /coderabbit-queue rather than upgrading — roughly every 70 minutes, one request goes out, and if the service replies that the limit was reached, the skill waits and retries, using the service's own stated wait time when it gives one. The everyday checks then accumulate into the same pattern: Opus for regular work and Fable for reviewing high-impact changes, while log reading runs on a lighter model to stay inside the Pro plan's ceiling.

Several habits came directly from mistakes, according to the account. After CodeRabbit reviewed a commit and the pull request was then changed, the display read "reviewed" while the newest changes had not been seen; he nearly read an old review as "no findings, OK," so skills now verify that the reviewed commit ID matches the latest one. A log analysis once read timestamps shifted by nine hours from confusing Japan time with UTC, and log-analyzer has specified UTC ever since. He describes /script-check as growing one check at a time from pitfalls he actually hit, such as writing a full-width character right after a shell variable and having it read as part of the variable name.

The division of labour is the piece he says changed the most. Inquiry and preparation move fast in Claude's hands, while "merge now" and "ship to production now" stay his calls; the approval table, keyed to a 40-character commit ID (SHA), exists so that a merge fails if anything changed after his sign-off. He notes that because this is self-funded development, he does not use GitHub's paid features such as branch protection, so the double safety net is his way of covering what the plans do not.

FAQ
What is the difference between a skill and an agent in Claude Code?
The author compares a skill (スキル) to a work manual — a written procedure called by name, like /status, that Claude may also invoke on its own. An agent (エージェント) is like delegating to a specialist: a separate Claude is given a job, such as reading logs and summarizing errors, and only the result comes back.
How does the author decide what Claude may do without asking?
He writes three tiers into the project's CLAUDE.md file: read-only tools like /status, /merge-ready and log-analyzer run on Claude's own judgment; write or cost-incurring ones like /pr-flow and /double-review need a proposal first; and merging, production deploys and full test runs need his approval.
What safety measures are used besides written rules?
He adds mechanisms separate from the rules: a Claude Code hook that checks commands before execution, a Git-side hook in case the first check is bypassed, and a dedicated script as the only route to production changes. The rules also say that if a safety mechanism stops Claude, it must ask him rather than rephrase to get around it.

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleGoogle's Playground builds games in minutes, Wired finds slop