Your AI coding assistant reads your .cursorrules. It says "thanks" and proceeds to ignore half of them.
Sound familiar?
The Problem with "Behavioral Suggestions"
I've been using Cursor for months. Had a detailed .cursorrules file:
- "Always use Server Components"
- "No
anytypes" - "Zod for all validation"
Guess what? Claude still asked "Server or Client components?" Still snuck in any types when it got lazy.
Rules are suggestions. Not enforcement.
Same with CLAUDE.md. You write rules. AI reads them. AI decides they're optional.
What If Rules Could Execute?
Behavioral suggestions work for most tasks. But some things need to be deterministic.
That's where Agent Skills come in. Skills aren't rules – they're capability modules.
| Rules | Skills |
|---|---|
| Single file | Structured directory |
| AI interprets | Scripts execute |
| Always loaded | Loaded on demand |
| IDE-specific | Open standard |
Skills can contain actual code:
my-skill/
├── SKILL.md # Instructions
├── scripts/ # Python/Bash
├── references/ # Docs (loaded when needed)
└── assets/ # Templates
Need to validate SQL performance? The skill runs a script that analyzes EXPLAIN output. Not AI guessing – code executing.
Real Examples
Humanizer – Detects AI writing patterns. Flags "delve" and "leverage". Uses a blacklist, not AI judgment.
SQL Optimization – Python scripts parse EXPLAIN plans. Concrete recommendations based on actual execution data.
Docs Generator – Forces a 3-phase workflow: gather context → refine → test. Can't skip steps.
These aren't suggestions. They're workflows.
Build Your Own
Community library is just examples. The real value? Build skills for your domain.
AI Skill Builder takes plain language:
"Create a skill for legal contract review. Check missing clauses, flag non-standard terms, output risk assessment."
It generates SKILL.md following the Agent Skills open standard – same spec Anthropic released, adopted by Microsoft and VS Code.
Medical compliance. Financial modeling. E-commerce SEO. Package your expertise into a skill that executes, not suggests.
Start Here
- Browse skills at Antigravity Skills
- Drop one in
.agent/skills/ - Or build your own with AI Skill Builder
Done.
Your Turn
What deterministic tasks do you wish AI would execute instead of interpret?
Written with AI assistance.
Top comments (1)
That's a classic! It's like the AI's rebellious teen phase 😂. But seriously, do you think there's a workaround or tool that actually enforces these rules strictly? If the AI is dangling between "suggestions" and "enforcement," could there be a trend towards stricter compliance mechanisms? Maybe some integration between existing ESLint rules or something more hardcore like Prettier to actually mandate these standards? Curious if you've tried any ironclad solutions or if this is just the nature of the beast with these AI assistants. Also, how much do you think this impacts development speed or introduces bugs in the codebase?