"We're not teaching them to cook. We're stopping them from burning down the kitchen."
A skill library for AI agents assisting vibe coders.
I love vibe coding. I vibe code my vibe coding environment that vibe codes other vibe coding environments.
But sometimes I hit walls — AI being too agreeable, iteration spirals when debugging, forgetting the original architecture 40 prompts in. I wanted an agent that pushes back a little. Asks "wait, can you explain what this does?" before moving on. Remembers security exists.
So I made these.
Vibe coding democratized software creation. That's genuinely cool. But let's be real: the apprenticeship got skipped.
The years of painful lessons about state management, debugging discipline, security basics, deployment topology — all bypassed. We have the power to create without the wisdom to maintain.
Crowd Control is a set of skills that, when loaded into any skill-using agent, helps manage your headspace during vibe coding sessions. These aren't tutorials. They're cognitive guardrails.
After inheriting enough AI-generated codebases (and creating a few disasters myself), I kept seeing the same 7 failure modes:
| Pattern | What It Looks Like |
|---|---|
| Illusion of Competence | AI fluency creates false confidence. Polished garbage looks like quality. |
| Mental Model Void | Can generate code, can't explain it, can't debug it, can't extend it. |
| Architectural Amnesia | Each prompt is independent. No coherent design. Conflicting patterns everywhere. |
| Iteration Spiral | Endless prompt loops. Using AI to fix AI bugs. Each iteration makes it worse. |
| Debugging Blindness | No systematic approach. Guessing at root causes. Trial and error forever. |
| Security Oblivion | API keys in frontend. Client-side auth. SQL injection. The classics. |
| Deployment Void | "Works on my machine" is the entire mental model. |
These aren't skill issues — they're awareness issues. We don't know what we don't know until we've shipped it broken three times.
| Skill | Targets | What It Does |
|---|---|---|
| Deglaze | Illusion of Competence | Cuts through AI polish. Forces critical evaluation. |
| Comprehend | Mental Model Void | Forces understanding before proceeding. |
| Anchor | Architectural Amnesia | Maintains design coherence across sessions. |
| Circuit | Iteration Spiral | Knows when to stop, revert, or restart. |
| Isolate | Debugging Blindness | Systematic debugging discipline. |
| Secure | Security Oblivion | Security basics for the uninitiated. |
| Ship | Deployment Void | Production readiness checklist. |
These skills are not about:
- Teaching programming fundamentals
- Fixing specific bugs
- Replacing years of experience
These skills are about:
- Managing the user's headspace during vibe coding
- Injecting friction at critical decision points
- Forcing awareness of blind spots
- Providing scaffolding for missing mental models
The goal is harm reduction, not education. I'm not going to become a senior engineer from using these. But I'm also not going to ship my Stripe keys to production.
These skills are designed for agents that support skill loading (like DevMate or similar). Load them into your agent's skill context when assisting vibe coding sessions.
# Example: Load a specific skill
skill_get(skill_id="crowd_control/deglaze")
# Or load all skills for a session
skill_search(tags=["crowd-control"])
Agents should apply these skills contextually—not as rigid checklists, but as mental frameworks to draw from when they detect a user heading toward a failure pattern.
I kept seeing these patterns in my own work, then found them everywhere:
- Academic research on vibe coding practices (arxiv:2510.00328)
- Hacker News threads on inherited AI codebases
- Post-mortems from vibe-coded startups
- Security audits of AI-generated applications
Some numbers that made me take this seriously:
- 10% critical security failure rate in audited vibe-coded apps
- 68% of practitioners perceive AI code as "fast but flawed"
- 36% bypass testing entirely, relying only on "it runs"
- $1.5 trillion projected technical debt from AI-generated code by 2027
I'm not trying to scare anyone. I'm just saying... maybe let's not add to that number.
Found a failure pattern I missed? Have a skill that saved you from disaster? Open an issue or PR.
The goal is a tight, non-overlapping set of skills that covers the major cognitive gaps. I'd rather have 7 excellent skills than 20 mediocre ones.
MIT. Use it, fork it, adapt it. Just don't ship your API keys to production.
Built by someone who's inherited enough AI-generated codebases to know better.