By skyforbes I'm relatively new to agentic coding and have been bouncing between Gemini, Antigravity, Claude, Roo, and Droid agents, but I'm hitting the same wall: they ignore rules, dump files in root, vomit mock data everywhere, and are little shits who do whatever they want ignoring whatever they don't feel like doing.
My current stack (which feels almost right but still missing something):
- BMAD for initial PRD generation
- Task-Master-AI to decompose tasks until they're 95%+ achievable
- Autopilot (the hidden MCP inside Task-Master) for TDD workflow
- rules/AGENTS.md for orchestration (just discovered tool-specific dotfiles like
.geminiactually get read—game changer)
Specific pain points I'm trying to solve:
- Rule amnesia: Agents forget project structure and conventions mid-session
- Root directory pollution: Random docs and mock files appearing everywhere
- Mock data obsession: Inventing fake APIs instead of using existing ones
- YouTubes who claim their tools solve hallucinations but my experience is 3 hours of wrestling for every 5-minute demo I try to recreate.
What I'm looking for:
- Guardrail that work in practice
- Validation tools, methods, scripts or hooks etc you've used to force compliance
- How to close the loop between task decomposition → test writing → code generation without a HitL
- The "connective tissue" between these tools that makes a chain more deterministic
I see people one-shotting whole games SaaS Websites, and even an OS in the browser but I'm trying to create a real, complex app and the thought of even 1000-shot doesn't seem likely. I know there's no magic bullet, but there has to be a better way than my profanity laden tirades that scare my neighbors or throwing my mouse at the window AGAIN. What are the undocumented tricks that actually work?
Yeah, I'm self-taught and probably missing something face-palm obvious—but I really don't what I should be asking as I don't know what I don't know!