Abnormal’s engineering vision is clear: AI should write the first draft of code, and humans should act as architects. But getting there requires tools that make AI-initiated development easy, fast, and reliable.

That’s the motivation behind NoraPR v2: a major upgrade led by Shrivu Shankar to unify and streamline Abnormal’s AI coding experience.

Before NoraPR v2, engineers faced several friction points:

1. Confusing tool choices: There were two primary tools: NoraPR, generated PRs via a background CI server; and Nora Code, ran code generation locally. The naming overlap and behavioral differences caused confusion and slowed adoption.

2. Slow background execution: The CI engine behind NoraPR could take 30+ minutes for simple changes, long enough for engineers to just write the code themselves.

3. Micromanagement traps: Using Cursor or Cloud Code locally allowed engineers to “hover” over the agent, nudging it through each line of code. This meant heavy human-in-the-loop, which resulted in only marginal productivity gains (~10% instead of 100%+).

4. Analytics gaps: Local workflows didn’t feed into AI-initiated metrics pipelines, so engineers weren’t getting credit for AI-driven work, and the org lacked visibility into adoption.

Combined, these issues slowed down the broader transition to AI-driven development.

Shrivu’s answer was bold and simple: Merge everything into NoraPR and add a “local” execution flag.

Now, the workflow is crystal clear:

1. Type nora-pr "your prompt"

2. Nora writes the code (locally or via CI)

3. A PR appears: tested, validated, and ready for review
   

This creates one path and one mental model for AI-initiated coding.

Key improvements include:

• Unified tool: No more NoraPR vs. Nora Code. Just NoraPR.

• Local mode:

  • Enables fast code iteration.

  • Still enforces best practices like front-loading context.

  • Ensures the agent, not the developer, does the writing.

• Background quality checks: Even for local runs, NoraPR applies automated testing and sanity checks before surfacing changes.

• Metrics integration: All NoraPR usage (local or CI) feeds into analytics pipelines to track adoption and quality.
  

The result? A consistent experience that trains engineers to architect, not micromanage.

In the last month alone, AI-initiated code changes dwarfed human-written changes. This is a major reversal from just three months prior.

With NoraPR v2:

• Engineers get out of the way of the agent.

• Nora writes full PRs automatically.

• The org unlocks triple-digit productivity gains, the kind that only come from fully autonomous workflows.

As Shrivu puts it: “I used NoraPR itself to build the new local mode and other changes. It took me 15 minutes. Building these slides took longer.” That’s the kind of speed Abnormal is aiming for.

NoraPR v2 is already reshaping engineering workflows:

• Faster iteration cycles

• High-quality, well-tested PRs generated automatically

• Consistent context handling ensures agents have what they need

• Less micromanagement, more real AI autonomy

• Better tracking and insights for AI-driven initiatives
  

It’s not just a tool. It’s a foundational shift in how Abnormal engineers write, review, and ship code.

With the unified NoraPR foundation in place, the next steps include:

• Expanding support for more complex, multi-file codebases

• Integrating richer architectural guidance into prompts

• Continuing to accelerate CI checks behind the scenes

• Standardizing development workflows org-wide around AI-first practices
  

The end-state? AI writes the code. Humans craft the intent. NoraPR is the bridge.

NoraPR v2 is a perfect example of Abnormal’s engineering culture:

• Start with a frustrating workflow

• Rethink it from first principles

• Automate everything possible

• Use AI to build the tooling for AI
  

Shrivu’s work didn’t just improve a tool. It moved the entire company closer to a future where engineering velocity is limited only by imagination.