# AI Engineering Training Focus — Northcove

Synthesized from meeting notes with Nina (full stack), Ben (Senior SWE, Business group), and Raj (Engineering org overview). Company: Northcove. Stack: Python backend, TypeScript frontend, monolithic app, ~100 engineers across Consumer/Business/Platform silos.

## Top Training Priorities (ranked)

### 1. Context Engineering & Company-Specific Knowledge (highest priority)
Every engineer cited context as the #1 blocker to effective AI use.

- **The problem:** Documentation is scattered and stale (Google Docs, Confluence, Slack threads). There's no structured way to feed documentation as context to Cursor. AI is "not good at Northcove-specific implementations." Ben's core principle: "Context is king — don't just say 'do the thing', tag relevant files upfront."
- **Train on:**
  - How to build and maintain a company-specific context library (reusable prompts, rules, `.cursorrules`/equivalent, tagging relevant files)
  - Retrieval-augmented generation (RAG) patterns for internal docs: indexing Confluence/Slack/Google Docs so AI can ground answers in current Northcove standards
  - Writing effective system prompts that encode Northcove coding standards, frameworks, and architectural conventions
  - Using similar examples from the codebase rather than general prompts (Nina: better results with concrete examples)

### 2. Evaluation, Testing & Guardrails for AI-Generated Code
- **The problem:** AI code "looks plausible but may introduce subtle issues." Nina lacks robust end-to-end and integration testing and monitoring. Her document-interpretation eval is labor-intensive manual labeling (100–250 docs) with non-deterministic results from small prompt changes. Ben avoids AI for production features because of hidden side effects and coupling. Raj is skeptical of AI for DB migrations and regulated-data work.
- **Train on:**
  - Building ground-truth datasets and eval harnesses for LLM features (proxy metrics for scale, labeled samples for accuracy)
  - Guardrails for AI-generated code: review checklists especially for database/regulated-data changes
  - Integration/e2e testing strategies to catch subtle AI-introduced regressions
  - Monitoring and validation systems for AI-driven features in production
  - When *not* to use AI: DB migrations, regulated data, significant non-trivial features

### 3. Prompt Engineering Standardization & Optimization
- **The problem:** Each engineer writes ad-hoc prompts. Ben wants organizational buy-in for rules, commands, and reusable prompts. Nina struggles to create comprehensive system prompts. Ben is aware of DSPy for prompt optimization but hasn't implemented it.
- **Train on:**
  - Standardized, Northcove-specific prompt templates and reusable commands
  - Prompt iteration methodology (small changes → non-deterministic outcomes; measure before/after)
  - Introduction to automated prompt optimization (DSPy-style) for teams with labeled data
  - "Directionally correct is the name of the game" — knowing when approximate AI output is acceptable (internal scripts) vs. when it isn't (production, regulated)

### 4. AI-First Workflow & Agent Adoption
- **The problem:** Most of the eng org is still "AI-second" rather than AI-first. Ben is a power user; Nina is a basic user. Agents/background processes are under-adopted. Everyone wants to delegate larger features to independent agents and reduce back-and-forth.
- **Train on:**
  - AI-first integration: using AI for questions, reproduction, documentation, planning, debugging — not just coding
  - Cursor power-user patterns: smart autocomplete, chat-for-planning-then-building, code search for structure
  - Background agents and subagents: test coverage, feature flag removal, debugging, long-running processes
  - Prompting for incremental changes (not just generating from scratch)
  - Moving from "editor" to effective AI collaborator without over-reliance

### 5. AI-Assisted Prototyping & Planning
- **The problem:** Tech spec documents try to solve strategic alignment, deadlines, resourcing, sequencing, and technical alignment all at once — too much, hard to review. Planning isn't prescriptive.
- **Train on:**
  - Using AI-generated prototypes (Vercel v0-style) for alignment — directionally correct, cheap to throw away
  - Splitting planning artifacts: separate strategic alignment from technical implementation details
  - Rapid prototyping for user research and PM/engineer alignment

### 6. Compliance, Data Isolation & Regulated Data
- **The problem:** A prior third-party data-sharing incident required reworking paid/organic surfaces. Litigation concerns make the team cautious. Nina uses Anthropic with organizational data isolation contracts. Raj flags regulated-data exposure as a key risk.
- **Train on:**
  - Data isolation contracts and what can/cannot be sent to AI models
  - Handling regulated/PII data in AI workflows
  - Vendor/model selection with compliance in mind

### 7. Cross-Functional AI Adoption (enablement)
- **The problem:** Non-engineering functions (PMs, BI) are adopting AI via Cursor Slack integration but lack foundation. BI hookup needed. User research via AI is "not ready" (only talking to 3 people, not enough volume).
- **Train on:**
  - Foundational AI literacy for non-engineers so they can benefit safely
  - Guidelines for PMs using AI for copy, change requests, and prototyping
  - Where AI fits (and doesn't) in user research and BI

## Suggested Training Structure

| Module | Audience | Format | Focus |
|--------|----------|--------|-------|
| AI-First Foundations | All engineers | Workshop + hands-on | Cursor power use, context tagging, "context is king" |
| Context & RAG for Northcove | All engineers | Lab | Indexing internal docs, custom rules/system prompts |
| Eval & Guardrails | Backend + AI feature owners | Lab | Labeled datasets, proxy metrics, review checklists, monitoring |
| Agents & Background Automation | Senior/intermediate | Workshop | Subagents, background tasks, incremental changes |
| AI Prototyping & Planning | PMs + tech leads | Workshop | v0-style mockups, split planning artifacts |
| Compliance & Regulated Data | All (mandatory) | Short session | Data isolation, regulated data, incident lessons |
| AI for Non-Engineers | PMs, BI, ops | Intro session | Safe AI use, Cursor Slack, copy/change requests |

## Success Metrics to Track
- % of engineers using AI-first (vs AI-second) workflow
- Reduction in time from PRD to deploy for bets (currently 1 week)
- AI-generated code caught by review/guardrails before production
- Eval coverage for LLM-powered features (vs manual labeling only)
- Reduction in "context not found" / stale-document incidents

## Assumptions Made
- Northcove uses Cursor as its primary AI IDE (confirmed across all three interviews).
- Anthropic is the preferred model provider (per Nina's document system) with data isolation contracts in place.
- Training should be role-differentiated: basic users (Nina) need foundations, power users (Ben) need advanced agent/DSPy content, org leads (Raj) need compliance and knowledge-management focus.