VibeSafe

AI-Assisted Development with Intent

Neil D. Lawrence

Trent.AI

The AI Development Paradox

  • You’ve used GitHub Copilot, Cursor, Claude Code…
  • AI writes code faster than ever
  • But have you debugged AI-generated code at 2am?

The Cost Model Has Inverted

  • Traditional development:
    • Writing code: Expensive (human time)
    • Documentation: “We’ll do it later”
  • AI-assisted development:
    • Generating documentation: Cheap (AI does it)
    • Debugging misimplementation: Expensive (discovery lag)

A Natural Reaction

  • “This looks like a lot of paperwork”
  • And you’d be right to think that…
  • …if we were still in the old cost model

VibeSafe’s Philosophy

  • Make intent explicit before implementation
  • Catch misunderstandings early
  • Cost: editing markdown vs unwinding code

Human-AI Collaboration

  • AI systems: Powerful but literal
  • They follow instructions precisely
  • Missing: shared context and intent

From Intent to Documentation

\[\text{WHY} \rightarrow \text{WHAT} \rightarrow \text{HOW} \rightarrow \text{DO} \rightarrow \text{DOCUMENT}\] \[\text{Tenets} \rightarrow \text{Requirements} \rightarrow \text{CIPs} \rightarrow \text{Backlog} \rightarrow \text{Doc Compression}\]

Core Components

Tenets: Your Project’s Principles

  • 5-7 guiding principles
  • Not rigid rules — principles to balance
  • Example: “User Autonomy Over Prescription”

Requirements: What, Not How

  • “Users can install with one command”
  • Not: “Create install-minimal.sh script”
  • Outcomes, not methods

CIPs: Design Before Implementation

  • Code/Capability Improvement Plans
  • Document design rationale
  • Review before implementing

Backlog: Execution Tasks

  • Specific implementation tasks
  • Created when CIP is accepted
  • Not when it’s proposed

Everything is Markdown + YAML

  • Standard file formats
  • Works with any AI assistant
  • Cursor, Copilot, Claude Code, Codex…

The Workflow in Practice

Example: Adding Authentication

  1. Check tenets: Which principles apply?
  2. Write requirement: “Users must authenticate securely”
  3. Create CIP: JWT vs session-based approach

AI Natural Breakpoints

  • After CIP creation → Review design
  • After acceptance → Create tasks?
  • After implementation → Validate

The What’s Next Script

  • Project status at a glance
  • For humans and AI
  • “What should I work on?”

Documentation Compression

  • After CIP is closed: compress into docs
  • Development history → permanent reference
  • Future developers understand what, not just how

Benefits and Trade-offs

What You Gain

  • Catch misunderstandings early
  • Shared context across team
  • Better onboarding (humans and AI)

What It Costs

  • Upfront documentation time
  • Learning curve for team
  • More files to maintain

When Does This Make Sense?

  • Working with AI assistants
  • Complex systems with multiple engineers
  • Long-lived codebases

The Real Question

  • Does this match how you actually work?
  • Or how you want to work?
  • What would you change?

Discussion

Questions for You

  • Have you experienced “AI misunderstanding” bugs?
  • How do you currently ensure shared intent?
  • What would make this more useful?

Open Questions

  • Right granularity for CIPs?
  • Integration with existing tools?
  • How to handle legacy code?

Try It Yourself

  • One-line installation
  • GitHub: lawrennd/vibesafe
  • Works with any AI assistant

Questions?

  • Let’s discuss
  • Your insights will help shape this
  • Thank you!

Thanks!

References