Supervise Claude closely? I catch everything.
Let it run free for a few hours? I discover the damage later.
Let it run free for a day? I'm refactoring for a week.
Here's the thing: 𝐢𝐭 𝐝𝐨𝐞𝐬𝐧'𝐭 𝐦𝐚𝐭𝐭𝐞𝐫 𝐡𝐨𝐰 𝐠𝐨𝐨𝐝 𝐂𝐥𝐚𝐮𝐝𝐞 𝐢𝐬.
It will slip. One shortcut. One forgotten convention. One "minor" inconsistency. You won't notice. Not immediately.
But the snowball has started.
- Day 1: a layer boundary crossed.
- Day 5: the shortcut becoming a pattern.
- Day 15: Claude learning from your codebase. Including the mistakes.
Technical debt doesn't announce itself. It compounds silently.
𝐂𝐥𝐚𝐮𝐝𝐞 𝐢𝐬𝐧'𝐭 𝐭𝐡𝐞 𝐩𝐫𝐨𝐛𝐥𝐞𝐦. 𝐂𝐨𝐧𝐭𝐞𝐱𝐭 𝐨𝐯𝐞𝐫𝐥𝐨𝐚𝐝 𝐢𝐬.
Claude knows the rules. I spent weeks refining them: architecture, patterns, conventions. All documented.
But sometimes it forgets. Just like us.
That's why humans do code reviews.
𝐒𝐚𝐦𝐞 𝐥𝐨𝐠𝐢𝐜 𝐚𝐩𝐩𝐥𝐢𝐞𝐬 𝐭𝐨 𝐀𝐈.
So I built a code-review agent that runs before every commit.
Here's what it does:
- Reads our architectural standards from a single CLAUDE.md file
- Checks every git diff against those rules
- Catches what Claude (and I) miss
Simple idea. Immediate impact.
I still do manual reviews. But now Claude catches most mistakes before I even look.
In 3 weeks:
- 𝐙𝐞𝐫𝐨 architectural violations shipped
- Review time cut by ~40%
𝐓𝐡𝐞 𝐥𝐞𝐬𝐬𝐨𝐧: The rules were already in CLAUDE.md. Claude knew them. But it will eventually forget. Skip one. Misapply another.
Having rules isn't enough. You need guardrails in your workflow.
Still iterating. Still learning.
How many slips before your snowball starts? ❄️
What I noticed as well is that Claude is over-influenced by the existing codebase. Whenever you introduce a new concept, library, framework it really takes a lot to make it deviate from the existing legacy code and stay consistent. Can your agent help with that ?