A lightweight governance layer that helps agents plan before coding

Feature request for product/service

– Other –

Describe the request

Hi everyone,

I’m not a professional developer — I’m a beginner who started using AI agents
to build a real project from scratch. Along the way, I kept running into the
same problems over and over: agents would jump straight into coding without
understanding what I actually wanted, leave no plan behind, claim things were
“done” without evidence, and create chaos that was hard to untangle in the
next session.

So I started writing down rules — just markdown files at first — to force the
agent to ask me questions before writing code, create a plan I could review,
and track what was actually done vs. what was just edited. Over time, these
notes grew into a small governance framework that I eventually extracted from
my main project into its own standalone repo:

It already has native Cursor support (via .cursorrules) and works with 10+
agents. What surprised me most was that even the cheapest, lowest-tier agent
models became noticeably more productive and coherent when working inside this
governance structure. The “interview first, plan before code” discipline seems
to compensate for a lot of model weakness.

I want to be honest about the trade-offs — this approach intentionally adds
friction. It’s bureaucratic by design. It uses more tokens because the agent
reads governance files and writes plans before touching code. These are
conscious choices: the idea is that a little overhead early prevents a lot of
chaos later. But I understand this won’t appeal to everyone, especially people
who want fast, zero-ceremony prototyping.

The core ideas that might be interesting for Cursor:

  • Phase-0 Interview: Agent asks structured questions one-by-one before
    writing any code — locks down platform, scale, and architecture decisions
  • Plan-before-code lock: No implementation until a reviewable plan exists
  • Honesty Boundary: Agent must distinguish between “edited” and “validated”
    — no more confident-sounding “done” claims without proof
  • Session memory: 4-pillar collective memory files that persist lessons,
    architecture, environment, and user preferences across sessions
  • Multi-Role Review: Every change reviewed from multiple perspectives
    (beginner, expert, accessibility, security)

Everything is MIT-licensed. I’m not pitching a product — I’m sharing an idea
that helped me as a complete beginner get real value from agents. If any of
these concepts are useful for Cursor’s roadmap, I’d be happy to discuss.

Thanks for reading.