Nightly AI Code Reviews That Run While You Sleep

2026-05-10

There's a window between writing code and reviewing it where problems hide. The faster you ship with AI, the wider that window gets.

Nightly AI reviews close it. You commit during the day. At 5 AM, Claude reviews everything that changed and drops a dated report on your project. By breakfast, you know what looked off.

Here's how to set one up on YokeDev.

What runs

A Claude routine triggers at 5 AM on your schedule. It connects to your YokeDev project through MCP. It calls the platform's built-in AI review, which:

Figures out which files changed since the last review.
Walks every check you've defined in .yokedev/aichecks/.
Writes one consolidated report at .yokedev/aichecks/reports/ai-review-YYYY-MM-DD.md.

The first run reviews your full codebase as a baseline. After that, each run scopes to the diff. A quiet weekend produces a one-line report. A busy Monday produces a longer one. You never re-pay the cost of reviewing the same line twice.

What it isn't

This isn't a replacement for code review, and it isn't a substitute for your git history. Your Gitea instance already tells you what changed and who. The routine tells you whether what changed meets the rules you wrote down.

Those are different artifacts. One is the record of work; the other is the assessment of it. The routine complements your existing workflow rather than competing with it.

Read-only by design

A scheduled AI session that can edit code is a footgun. One ambiguous prompt and your codebase mutates while you sleep.

The routine prompt is calibrated to refuse modifications. The only write it permits is the final report file. No commits. No edits. No package installs. Your code doesn't change unless you change it.

What the report looks like

Each check returns one of four states with file:line citations where it matters:

PASS: check found nothing to flag.
FAIL: specific lines violate the check, with paths and line numbers.
PARTIAL: check couldn't fully evaluate.
N/A: check didn't apply to this codebase.

Read it. Fix what's real, tune what's noise, ignore what's intentional. Three triage paths, every morning.

Setup is one paste

In claude.ai/code/routines, create a routine. Name it. Paste the prompt body (full text in the YokeDev docs that ship with your project). Enable your YokeDev MCP connector. Schedule it daily at 5 AM. Click Create.

Click Run now to verify the wiring. From then on, the routine fires on its own.

Total setup time: about five minutes.

Beyond nightly

The same routine, no rewrite needed, also works for:

Pre-migration reviews: click Run now the day before a risky schema change to capture a baseline.
Customer hand-offs: run one final review before transferring a project. Hand the report along with the code.

Same prompt body. Different reasons to fire it.

Try YokeDev

YokeDev gives every project a dedicated VM with an MCP server, a deployment pipeline, and the AI review tooling pre-installed. Connect it to Claude in two minutes; you're a Run-now click away from the first report.