What AI Can't Do
Set the right expectations for AI coding tools by understanding what they're great at and where they fall short
“AI Can Build Anything” Is Half True
AI coding tools can do a remarkable amount. But they’re not magic.
When expectations are off, you end up frustrated — “why doesn’t this work?” or “AI is useless.” Knowing the boundaries upfront helps you work with AI much more effectively.
What AI Is Great At
- Generating standard code — forms, CRUD operations, layouts
- Finding bug causes — paste an error message and it explains what’s wrong
- Applying existing patterns — “make another page with the same structure as this one”
- Writing documentation — reading code and producing explanations
- Refactoring — restructuring working code into something better
- Writing tests — generating test code for existing functions
What AI Struggles With
1. Deciding What to Build
AI is great at “how to build it,” but it can’t decide “what should be built.”
# AI can't do this
"Come up with a service that users will love"
# AI can do this
"Build a TODO app where users can add, complete, and delete tasks"
Requirements are your job.
2. Visual Design Taste
AI can produce UI code, but “looking good” and “feeling right” are subjective.
- Color palette harmony
- Spacing and balance
- UX that feels intuitive
AI can give you a starting point, but the final call is yours.
3. Reading the Room
AI resets with every conversation (CLAUDE.md helps, but has limits).
- Unwritten team conventions
- “We tried that approach before and it failed”
- Business priorities and trade-offs
This is exactly why writing a CLAUDE.md matters.
4. Sustained Consistency Over Long Sessions
When you try to build a large system in a single conversation, AI can lose track — contradicting earlier decisions or drifting from the original plan.
Fix: Work in small chunks. Don’t ask for everything at once.
5. Final Security Judgment
AI knows security best practices, but:
- It doesn’t fully understand your actual data flow
- It sometimes generates “probably fine” code with full confidence
- Authentication and authorization in production must be verified by humans
6. Bleeding-Edge Information
AI has a knowledge cutoff date.
- A library version released yesterday
- An API spec that changed this morning
- The latest security vulnerability
Always check official documentation.
Healthy Expectations
| Frustrating mindset | Productive mindset |
|---|---|
| AI does everything for me | AI is a skilled assistant. I give the directions |
| Perfect code on the first try | It takes a few rounds to get it right |
| Errors = AI’s fault | Errors = a chance to debug together |
| AI should just know | I need to provide context every time |
Turning “Can’t” Into “Can”
Many things AI can’t do alone become possible with human + AI together.
- Design looks generic → Share a reference URL: “make it look like this”
- AI loses consistency → Write rules in CLAUDE.md
- Need current info → Paste the relevant docs: “based on this”
- Task too big → Break it into steps yourself, then delegate each one
AI is a tool. The people who understand how to use it get the most out of it.
Next Steps
- Writing Effective Prompts — Tips for giving better instructions
- Understanding AI Tool Permissions — Know what AI is trying to do