ainativeeng
search
48 skills. Prove them or don't.

AI-NATIVE ENGINEERING

You're shipping AI to production.

How many of these skills can you actually verify?

THE QUESTION
COMPETENT or weak?
THE BENCHMARK
NO VIBES. just criteria.
ENGINE_LOG: CORE_LOGIC
>> SELF-ASSESSMENT: 48 SKILLS
[ prompt-eng ] VERIFIED
[ rag ] WEAK
[ eval-driven-dev ] WEAK
[ guardrails ] VERIFIED
[ cost-optimization ] WEAK
[ observability ] VERIFIED
[ mcp-design ] WEAK
18/48 VERIFIED 30 GAPS FOUND
WHY_THIS_EXISTS

"AI engineer" is the fastest-growing job title of 2026 and nobody agrees what it means. Credentials don't prove anything. Self-assessment is noise.

This repo says — in plain language — what competent looks like at each of the 48 skills, and what weak looks like. Then it gives you the eval.

SAMPLE_WEAK_BARS

What "weak" actually looks like.

Every skill gets a one-line diagnosis. Pulled verbatim from the README.

WEAK · RAG

Dumps the top 50 chunks into context and prays.

WEAK · Evals

Ships to prod because "it felt right in testing."

WEAK · Guardrails

One regex blocklist for "bad words." Blocks "Scunthorpe."

WEAK · MCP Design

Shipped an MCP server with 30 tools named get_data_1 through get_data_30.

WEAK · Caching

Pays OpenAI to answer "what is your refund policy" 4,000 times a day.

WEAK · Streaming

Buffers the entire response server-side, then dumps 2KB of text into the UI at once.

COVERAGE

10 domains. 48 skills.

The full stack of what AI-native engineers ship. From foundations through governance, with nothing hand-waved.

Foundations
5 skills
Retrieval & Context
5 skills
Agents & Orchestration
6 skills
Evaluation & Quality
6 skills
Structured Output & Data
4 skills
Infrastructure & Operations
9 skills
Data & Fine-Tuning
3 skills
Product & Strategy
3 skills
Governance & Safety
4 skills
Advanced Patterns
3 skills
HOW_TO_USE_IT

Three ways in.

Subscribe for one skill unpacked per week.

Subscribe to the newsletter

By Ron Diver · rondiver@gmail.com