AI Writes Code But Misses Unwritten Rules: The Real Issue Is Human Management

01 The Spark: AI Fixes a Bug But Breaks the Architecture

Engineer tony shared an incident: an AI code review passed a PR, but a senior engineer later found the bug fix undermined the original architectural design. The team's formation was misaligned.

Engineer leo noted this happens often. Adding project-specific historical bug-fix skills aims to prevent the team from repeating known pitfalls.

Product manager sarah observed that by the time such lessons are documented, the model may have already evolved past those issues. She suggested documenting only project-specific idiosyncrasies and letting the model handle common problems.

This hit the mark: project-specific context is something AI can never fully learn on its own.

02 AI Has No Philosophy; Code Reads Like a Construction Site

Engineer carl offered a vivid metaphor: "AI can solve problems, but it has no philosophy. Reading good code used to feel like reading a poem; now it feels like watching a construction site. You can't say it's uninhabitable, but it's just not comfortable."

The analogy fits. The code runs, but something is off — like a bowl of noodles with all the right ingredients yet missing the essential flavor. That missing flavor is experience, style, and the team's unspoken rapport.

03 The Rules Debate: The Problem Isn't Rules, It's Unclear Rules

Architect witon raised a deeper issue: even with rules defined, AI may not follow them, and changing requirements leave AI confused about which rule applies.

Tony replied: "It's not a lack of rules; it's that the rules aren't clear enough."

Witon pressed: "How granular must rules be to be clear?"

Tony answered: "People have hidden motives; they keep unwritten rules, leaving a wide gap."

Unwritten rules — once that phrase surfaced, everyone knew where the real problem lay.

04 AI Doesn't Understand Unwritten Rules

Tony gave an example: layering conventions in his team were never written down; they were a mutable, tacit agreement. Current AI doesn't grasp such unwritten rules — like knowing to toast the host first at a banquet, or positioning the fish head toward the leader.

Witon added: "Humans share tacit understanding; a glance tells you what to do. AI has none of that."

AI is like a top-tier university graduate — brilliant, zero experience, outperforming other inexperienced peers. With a reliable lead, it performs well. But the question remains: who is leading?

05 Whack-a-Mole: Fix One Issue, Another Pops Up

Witon described a frustrating pattern: ask AI to change something, it fixes that but breaks something else. Tell it it's wrong, it corrects course, yet next time it invents new strange errors. Commit the fix to memory, that specific error stops, but others keep emerging.

Tony admitted: "It's still about the person wielding it. Honestly, most juniors I've mentored perform worse than AI."

He then compared AI to the legendary Red Hare horse: in the hands of Guan Yu or Lü Bu it's a peerless steed; in a foot soldier's hands it might end up as horse meat in a few days.

AI's ceiling isn't set by the model — it's set by the human using it.

06 Root Cause: It's Always a People Problem

The discussion converged on a single conclusion.

Witon said: "Tool software usually lacks business context, but many applications and B2B systems need that data. Humans can't enumerate every data rule."

Tony summarized: "The more standardized the project management, the more comprehensive the data accumulation, the more reliable AI becomes in principle. I've never encountered a problem that wasn't a people problem; every issue I've seen is a people problem."

It's all a people problem.

That sentence is the answer to the whole discussion.

AI doesn't know unwritten rules because we never wrote them down; AI plays whack-a-mole because we didn't curate the data; AI produces construction-site code because we never gave it a clear architectural blueprint.

Closing Thoughts

The insight isn't about AI's flaws. It's that AI acts like a mirror, reflecting our own management maturity.

Is your project management standardized enough? Is your data accumulation comprehensive enough? Have your unwritten rules been documented? If the answer is no, don't blame AI when things go wrong.

Blame yourself.