How I use AI tools, agents, and skills in my engineering work

AI tools are part of my engineering workflow. I’m interested in how they can help with debugging, code review, research, and validation while keeping the quality of the work visible and testable. That has led me to build my own agents and reusable skills around the tasks I repeat.

With GitHub Copilot and custom GitHub agents, I focus on giving the agent useful context. A stack trace is a starting point, but it rarely tells the whole story. The surrounding process, the business workflow, and the behavior the system is supposed to preserve all matter. My debugging agents are designed around that wider context so the investigation can reach the cause of a failure.

I also build reusable skills for code review and code validation. A skill gives a recurring task a consistent set of expectations: what to inspect, what evidence to collect, and how to report the result. Agents can select and combine those skills based on the change they are reviewing. This helps make the review process explicit instead of relying on a different prompt every time.

Another area I explore is multi-agent review. I designed an academic review loop that looks at the flow of an argument, APA 7 compliance, paragraph structure, clarity, and whether the evidence supports the conclusion. The feedback is organized as a training plan. I’m interested in how the same approach can help with design documents and pull requests, where several kinds of review are useful.

Research needs its own checks. My Research Accuracy Agent Harness structures source gathering, grounds claims in evidence, and verifies findings before reporting them. I also compare model outputs with defined scoring rubrics. A polished answer alone does not tell me whether the reasoning or evidence is sound.

Outside these focused workflows, I operate an OpenClaw-based personal assistant with task-specific bots for routine activities. Working with several agents in daily use gives me another perspective on context, coordination, and the boundaries of automation.

My approach is to define the task, give the tools enough context, make the output verifiable, and keep human review where the stakes require it. I want to use this journal to document how that approach evolves: the AI tools I try, the agents and skills I build, and what I learn through my PhD in Information Technology. Future posts will explore practical tool experiences and lessons from studying applied AI.