Use case B40 · Business
Have an automated check review your software projects each morning, handle the routine changes, and bring you only what needs a decision
- Operations
- Advanced
- Claude Code, GitHub, Claude Security plugin
- I use this
The short version
What I do
I have dozens of software projects, most of them built by AI agents. At 4 a.m. a job on my always-on Mac reviews every open pull request. Drafts that pass their checks get promoted, and drafts that sat for 7 days or more are closed with a note, branch kept. Then it writes one digest: what would merge on its own, what waits on me, what goes back to its author, and which projects it doesn’t cover. Its review comments lead with what the change does, whether it’s safe, and what I need to do, in plain English, because I’m an accountant, not an engineer.
Behind the digest is a merge bot. Each project gets a tier: some can merge on their own, some always ask me, and public ones are never touched. A change merges only when its tests pass in a locked-down sandbox and a reviewer that didn’t write the code passes it. Three things always come to me as a one-line question: accounting or tax judgment, anything public, and database changes that drop, rename or delete data. I’ve run the merge bot in dry run so far, where it reports what it would have merged instead of merging.
Now and then I also run a security scan over a whole app. On LedgerClaw, my bookkeeping app, I ran the Claude Security plugin, a separate panel checked every finding before I saw it, and agents opened a fix for each one with a test that fails on the old code. I reviewed the fixes and approved the merges.
What it gets me
I start the day with one list instead of opening dozens of projects. A typical morning: 2 would merge, 47 wait on me, 2 go back to the author, 4 projects not covered. The dry run taught me that the reviewer’s sign-off has to come from a separate bot account, or the bot can’t tell an independent pass from an agent passing its own work, and that anything opened more than 14 days ago should never merge on its own. The LedgerClaw scan found 21 issues (none critical or high, 13 medium, 8 low), all confirmed and all fixed the same day I approved them. One fix cut off my own connector tokens until I made new ones.
Prompt
I own several software projects and want a morning review digest in plain English. The open pull requests, check results and review records are [paste or connect]. Project rules are [which always need me and which are excluded]. My judgment areas are [such as accounting or tax logic].
1. Sort requests into four groups: needs my decision, would be ready to merge, goes back to the author, and not covered by adequate checks.
2. For each item needing me, state what changes, the concern and the one question I must answer.
3. Require passing checks on the current version and a review by someone who didn’t write it before marking anything ready.Always flag public changes and database changes that drop, rename or delete data. Give a short grouped digest, with missing evidence marked unknown. This is a dry run: don’t merge or close anything.
Do it yourself
These are suggested steps for you, not a copy of my exact setup.
What you need
- Your projects on GitHub with tests that run on each pull request.
- An AI assistant that can read GitHub, such as Claude with the GitHub connector or a coding agent with the GitHub command-line tool.
- For the full version, an always-on computer or a scheduled cloud job.
Start simple
- Once a day, ask your assistant to list your open pull requests and their check status.
- Paste the prompt above and read only the “needs my decision” group.
- Note what it got wrong and add those rules to the prompt.
Take it further
- Schedule it to run every morning before you wake up and save the digest where you’ll see it.
- Add an independent reviewer: a fresh AI session that reviews code it didn’t write and answers PASS, FAIL or ASK ME.
- Sort every project into tiers (merges on its own, always asks, never touched) and run a merge bot in dry run until you trust a week of its digests. Give the bot its own GitHub account.
- Run a security scan over each app now and then. Ask it to verify each finding, then fix them one small pull request at a time, each with a test that fails before the fix.
Check your work
- Spot-check the “ready to merge” group against GitHub for a week before you trust it.
- Make sure the checks it trusts ran on the exact commit being merged, not an earlier one.
- Before merging a fix that changes permissions or tokens, ask who will lose access.
- Watch the “not covered” group. Projects with no tests get no safety net from this.
- A clean security scan isn’t proof the app is safe. It means this pass found nothing more.
I use this. My own setup. I use it in my own work. Prompt adapted for this library, inspired by the case. Business size: One-person firm. Results are quoted as the source claimed them and never improved. Spot an error or a broken link? Email charlie@cbarmorecpa.com.