12 Unfiltered Observations on Working with AI

01 AI‑First Mindset

Before tackling any task, consider whether AI can help; this is both the easiest and hardest step because it requires forming a clear problem statement for the model.

Explaining what you want to the AI is rarely simple; the more complex the request, the more time spent on clarification.

02 Willingness Trumps Ability

AI can level the playing field: a highly motivated developer with limited skills can achieve results once AI assists, while a senior specialist who relies on personal knowledge may see their unique advantage diminish.

03 Outsiders Guiding Insiders

Non‑experts equipped with AI can sometimes out‑perform seasoned engineers; for example, feeding a k8s migration plan and context to an LLM revealed feasibility issues that the original proposal missed.

04 Judgment Is Still Essential

AI often lacks global consistency and can hallucinate; unchecked AI‑generated code or documentation can flood a system with low‑quality artifacts, so human review remains indispensable.

05 AI Efficiency Short‑falls (Head and Tail)

The "head" problem is articulating a clear request; the "tail" problem is rigorous acceptance testing and manual review after AI assistance, which can become bottlenecks.

06 Prompt Engineering Matters

Even as models improve, the choice of keywords (prompts) determines which internal parameters are activated; effective prompting is crucial for reliable outputs.

07 Don’t AI for AI’s Sake

Quantify efficiency gains instead of indulging in AI for its own sake; deterministic tasks should be solved with dedicated tools rather than repeatedly invoking large models.

08 AI May Exacerbate Overcapacity

AI can increase content volume without proportional business value, leading to homogenized output, potential layoffs, and a race to produce more without higher revenue.

09 Team‑Level Reuse Grows Bottom‑Up

Reusable assets such as documentation, skill guides, rules, and functional agents (e.g., CodeReviewer, TapdBugFixer) naturally accumulate from concrete scenarios and benefit the whole team.

10 Knowledge‑Base Focuses on the Knowledge Itself

Storing and retrieving knowledge is a generic problem best solved with existing services; building a custom solution requires extensive tuning and suffers from limited data for small teams.

11 Conclusion: Ability Can Be Supplemented, Willingness Cannot

When AI gets stronger, the remaining human edge is asking the right questions, maintaining judgment, and having the willingness to push boundaries.