Introduction: The more AI makes things easier, the wider the gap among engineers
Generative AI (LLMs) has become commonplace, and coding has accelerated. But at the same time, it has become a situation where anyone can produce something that looks right, and the value gap tends to appear in places other than the implementation itself.
This article distills the essential skills to become a strong engineer in the AI era into 10 items, framed so that you can imagine what to do tomorrow. We'll cover difficult topics too, but we'll keep it approachable.
10 Essential Skills: From Being a User of AI to Making AI Your Weapon
1. Problem Framing (Articulating what to solve)
AI can generate many candidate solutions. But what to solve is vague, and you’ll quickly lose your way.
- Goal: Revenue? Cost? User experience? Risk reduction?
- Constraints: deadlines, budgets, legal, existing systems, operations
- Success Metrics: SLO/SLI, CTR, churn, effort reduction, etc.
In practice, when you hear that AI could be used to do something, the strong engineer is the one who immediately asks, “and what would need to get better, and by how much, for it to be considered a success?”
2. Requirements Definition and Specification (Turning vague wishes into implementable specs)
Generative AI can create convincing text or code, but when requirements are vague, the output quality becomes vague. The strength of engineers is in drawing the boundary of the specification.
- User stories (e.g., who, for what purpose, what to do)
- Acceptance criteria (Given/When/Then)
- Non-functional requirements (performance, availability, audit logs, permissions)
Even when consulting with AI, having well-defined specifications increases the chance of a good output.
3. AI Literacy (Understanding the strengths and weaknesses of LLMs)
“If AI says it, it must be correct” is dangerous. Strong engineers treat LLMs as a convenient reasoning engine while also understanding their limits.
- Hallucination: can confidently tell a lie
- Data freshness: assumes it does not know the latest specs or internal circumstances
- Probabilistic outputs: same question yields different answers
Conversely, its strengths include summarization, translation, ideation, scaffolding code, and test-case ideas. If you can design where to rely on AI, and where humans verify, you’re a notch higher.
4. Instruction Design, Not Prompts (Building reproducible requests)
Reproducibility beats one-off prompts in real work. The recommended approach is to package: role, objective, constraints, input, output format, and evaluation criteria.
Example (key points): You are an SRE. Objective: prevent outage recurrence. Constraint: do not change the existing AWS configuration. Input: incident logs. Output: hypothesis of cause, additional investigation items, and proposed permanent countermeasures, ordered by priority.
With this alone, AI’s answers become much more usable as reference material.
5. Data Design & Data Quality (Winning with data even before AI)
The success of an AI product is often determined more by data quality than by the model. This is especially evident in RAG (retrieval-augmented generation).
- Logging design (event name, attributes, granularity, handling of missing values)
- Data normalization, master data management, data contracts
- PII handling, anonymization and masking
Tool examples include BigQuery/Snowflake, dbt, OpenMetadata/Amundsen, Great Expectations, which commonly appear in practice.
6. Verification Capability (Designing evaluation metrics, tests, and measurement)
In the AI era, it’s stronger to prove that it works correctly than merely to make something. Evaluating generative AI is particularly difficult; systematization is essential.
- Offline evaluation (accuracy/recall on known question sets)
- Online evaluation (A/B testing, gradual rollout, guardrails)
- Automated evaluation (LLM-as-a-judge) plus human review
Tool examples: LangSmith, Langfuse, Weights & Biases, Evidently, OpenTelemetry, etc. Start by aiming to be able to quantify quality with numbers.
7. Security / Privacy / Governance (The guardian wins in the end)
Incidents arising from AI deployment often involve information leakage via prompts or training data and lax permission design. Engineers who master this are highly valued.
- Handling of confidential information (internal policies, DLP)
- Access control (least privilege, audit logging, key management)
- Prompt injection countermeasures (input separation, allowlists, control of tool calls)
Additionally, regulatory topics like the EU AI Act are increasing. Even if you’re not a legal expert, noticing that a requirement may touch regulations or contracts is a strength.
8. System Design Capability (Drafting architectures that incorporate AI)
AI features alone don’t yield value; they must be properly integrated into existing workflows and systems.
- RAG (vector DB + search + generation)
- Agents (tool execution, workflow automation)
- Caching, rate limiting, fallbacks (designs that don’t break when AI is not available)
Vector DBs include Pinecone, Weaviate, Milvus, pgvector, etc. More important than tool names is the perspective: designing with awareness of where things can fail.
9. Cost & Performance Optimization (Balancing speed, cost, and safety)
LLMs are convenient, but misusing them can inflate costs. Performance gains in the AI era involve not only CPU and SQL but also tokens.
- Prompt shortening, context compression, and use of summarization
- Gradual fallback to smaller models (routing)
- Batch processing, caching, and streaming responses
As a practical rule, dashboard metrics like “cost per request,” “daily cap,” and “peak latency” should be visible to the team to drive awareness.
10. Communication & Writing Skills (In the AI era, your ability to convey ideas is a weapon)
Finally, this may seem mundane but is extremely effective. AI can draft, so those who can craft coherent text and secure agreements gain strength.
- Preserve design intent (ADR: Architecture Decision Record)
- Review-friendly PRs (short description of purpose, changes, and validation)
- Translate technical details into business value for stakeholders
Engineers who communicate quickly become increasingly valuable in the AI era.
Putting it into practice tomorrow: How to develop the 10 skills in the shortest path
Step 1: Break down your work with AI as a given
Break your current work into decisions (humans) and tasks (AI). For example, push implementation toward AI and shift yourself toward specification, testing, and risk assessment.
Step 2: Build small, and turn evaluation into a system
Don’t stop at a PoC; include evaluation metrics and logging from the start. If the AI only "works," that’s risky; only when it’s measurable does it become a product.
Step 3: Once a week, write an “AI-resilient” design
Recommended: every week, leave an ADR or design note. It’s fine to have AI draft it, but finish by polishing in your own words. This trains both writing and design skills at once.
Conclusion: Instead of fearing AI, increase the areas you can delegate
Engineers who are strong in the AI era don’t possess magical skills. They follow the thread of problem framing → specification → verification → operation and strategically expand the parts they can entrust to AI.
If you’re unsure where to start today, I recommend beginning with the “Evaluation (Skill 6)” and the “Documentation/Writing (Skill 10)” first. The outcomes are easy to observe and have a strong impact on the team, so you’ll feel growth soon.
