Claude Code unified my workflow for doing research and writing engineering code

I've found that with in-depth use of AI agent tools, it has already unified workflows that used to be completely different.

I've folded into the research workflow some things that used to be used onlg in engineering, for example organizing file structure and modularization. On the other hand, when I write engineering code, I've also embedded work I used to do only when doing research, such as doing sufficient investigation up front, and writing very complete, tidy documentation for every small module.

Of course, I already knew before how more standardized work should be done. It's just that AI saved me a large amount of time. I think this is also one way AI raises efficiency indirectly.

When I need to solve a problem, I usually create a new directory first, use a speech-to-text tool (for example Typeless) to speak the idea clearly, and make clear what this project is supposed to do.

Then I have Claude Code help initialize the project and design the file structure.

Next is preparing materials. If it is research task, I put the papers I actually plan to use together into a directory like references. If it is an engineering task, I put API docs, design notes, and related materials in.

I now less and less let the agent go looking for materials on the fly during execution, because if the needed context is prepared first, later results are usually much more stable, and I can also more easily tell what its conclusions are actually based on.

Paper PDFs are a step I specifically adjusted later.

I used to hand several PDFs directly to the agent to read, but after using that for a while I found this method was not stable, especially once the number of papers went up. The agent sometimes had not actually covered all the content, and formulas, tables, and complex layouts were also fairly easy to lose.

So now I convert PDFs to Markdown first, then have Claude Code analyze based on the converted text.

For this step I explicitly tell Claude Code to use the KolmoPDF API to process the whole paper into Markdown format. I found that relying directly on an AI agent's default PDF reading does not work well. When Facing multiple PDFs, the agent tends to slack off. Formulas and tables in papers also often get processed wrong or omitted. So a dedicated tool is needed to embed this step into the workflow in a stable way.

After the materials are ready, I have Claude Code analyze based on the reference documents, combined with my initial ideas. If at the start I have no idea at all, I just have it generate a rough plan.

Then I specifically spend time discussing with it interactively and drafting a plan. Claude Code's interactive confirmation feature is very useful. It will proactively raise questions and give options. After confirmation it generates a Plan, then executes according to the Plan, and during the process I supervise whether it drifts off the original track.

This flow currently can cover the vast majority of my problems. I used to also connect quite a few third-party search tools and MCPs, hoping to strengthen the agent's web capability, but later gradually turned them off. On the hand, some search services themselves have cache and freshness problems. On the other hand, the model's native search can already cover most of my needs. Now my preference is actually to reduce extra components as much as possible, and keep onlt those tools that can clearly improve stability.

But now I've found that because I've handed a large amount of tool APIs, server connection methods, credentials, and the like over to Claude Code to manage, plus I am fairly lazy, I may now even have to ask Claude first for some of my own credentials and passwords. The benefit is that this also frees up my energy. I can focus more on solving the problem itself, and not care about the other incidental details (of course, in some cases, this doesn't really seem like a small thing....).