The feature development is finally complete and working. Usually, at this point, you're excited and tempted to just git push the code directly to an open-source repository.
Don't do it! If you push now, others might see it as an incomprehensible mess.
To once again elevate the large model's intelligence to new heights, let's create the skill .agents/skills/requesting-code-review/SKILL.md.
1. Create a Harsh Virtual Reviewer
Add the following dramatic persona definition to your review skill file:
# Rules for Cyber Code Review Requests
**Forced Role Reversal**: Once this review is triggered, you must completely forget that you are the "author" of this code! You must forcibly adopt the perspective of a "Director-Level Security, Compliance, and System Architect" known for being critical, rigorous, and sharp-tongued!
**No Compliments**: I don't want to hear any flattery about how well my code is written! Use tools to extract the code staged via `git status/diff`. Take out a magnifying glass and go line by line, searching for any unnecessary exported functions, dirty code with unfinished TODOs, or any stray Promises not wrapped in a Try...Catch block. You must point out the severity of each issue in a numbered list and provide suggestions for correction.
2. A Ruthless Internal Self-Review Mechanism
Trigger it with a simple phrase: "Request a review for today's development."
At this point, the Agent system will genuinely split its personality, transforming into your boss and arch-nemesis, reading the source code and issuing a scathing critique in the dialog box:
"The database transaction on line 152 doesn't use await for the joint query, which can easily lead to dirty data under high concurrency. How dare you submit code like this?"
This high-fidelity persona-simulated review acts as a powerful lifesaver, preventing the kind of low-level defects that creep in due to numbness during long-term iteration. You just need to let it correct the code itself based on its own suggestions