--- title: 'Getting value out of coding agents for low-urgency, low-impact projects.' date: 2026-05-04T12:00Z tags: - using large language models - website has_math: false --- I've been a somewhat late adopter to coding agents: until ~1 month ago I was mainly using cursor and asking it to suggest local edits, or occasionally bigger refactors. In the past 1 month I've been trying to change that. In my work this has mostly taken the form of giving coding agents bigger open-ended tasks (e.g. "plan out this big refactor"). However, outside of work I've also tried just doing _more_ tasks. In this post I'll walk through some use cases where I used coding agents for low-urgency, low-impact tasks that I'd previously been procrastinating on. What's impressive is not the task itself, but how little mental attention completing the task required. **NOTE**: all these tasks were accomplished with Claude code, which I chose simply because I've had a personal subscription to Claude for the past ~1 year. ## Task 1: CV upgrade A few years ago I created a programmatic CV repo: a bunch of LateX snippets about individual items that could be combined together into a custom CV by putting a bunch of `\input` statements in a boilerplate file and setting tags with the `tagging` package. Unfortunately, it slowly grew into a tangled mess.[^mess] Since I'm not applying for jobs now I thought the fix wasn't urgent, so I let it be. [^mess]: The mess was essentially: 1. Files that were mainly supposed to contain _content_ also contained some formatting, so I couldn't really customize the format of any derived CVs. 2. As I created CVs of different lengths, I added ever more tags into the context files (e.g. `phd_short`, `phd_short_no_group_name_bullet_point`, etc), causing them to bloat. 3. Changes to content for 1 CV spilling over into other CVs (e.g. adding an extra line about my work at MSR caused my 1 page CV to expand to 2 pages). Last week I set Claude code on it. It felt like the perfect task because it was: - Low risk (worst case I revert the changes with git) - Low urgency (I won't need a CV until I apply for another job) - Safe place for code I don't understand: ultimately the repo is a build script for a document, so as long as the document looks good everything is fine (there are no "edge cases" to worry about like when you are writing an algorithm). After lots of planning and thinking, Claude recommended migrating to quarto, which would compile to pdf using typst. The format is simpler than Latex and there are apparently more CV formats available. It migrated my files, then wrote a few custom lua scripts to handle the compilation to typst. In the future I could also put its html output on my website. Ultimately I was pretty happy with the result. The decision on which backend to use ultimately took a fair bit of time, so altogether this migration maybe took ~5h of my time. This was probably more than it deserved, but a large part of this was learning about Claude itself and learning more about quarto (which I recently learned about in the context of my work anyway). ## Task 2: (not) migrating my wiki to Obsidian I've kept a wiki for personal notes since the beginning of my PhD. It uses the [vimwiki](https://github.com/vimwiki/vimwiki) vim plugin. However, I'd heard good things about [Obsidian](https://obsidian.md/) as an alternative wiki manager and was interested in exploring it. It is still based on markdown files stored locally (so I won't be locked into a third party), but apparently has a lot of other features. I opened Claude in my wiki and asked it to explore an Obsidian migration. Ultimately it told me that the experience of editing in vim was likely to be similar, that the editor might be _worse_ if I normally like vim keybindings, and it's mainly worth the migration if I plan to use its features like exploring backlink pages. I concluded that it wasn't worth the switch. This conversation took ~15m of my time (not counting time when I switch tasks while claude was thinking). That felt about right for what this task was worth to me. ## Task 3: fixing math rendering on this website I noticed an error where `\mathcal` letters on my blog wouldn't render correctly on Google Chrome. My interim solution was just to not use `\mathcal` fonts, which is fine because none of my posts are so technical that the notation and distinction really matters. However, every time I opened my website repo I saw the item sitting in my uncommitted `TODO.md` file. I asked Claude about the error, and it suggested migrating to math displayed with SVG instead. We got it working after 1 failed attempt, and it solved the issue. However, I wasn't a fan of all the math being images. So I asked it to think of a different solution, and it suggested rendering math with `KaTeX`. I'd tried `KaTeX` before but rejected it. I didn't totally remember why but I had a note last year saying it was rejected. It might have been due to some math code not rendering properly in a jupyter notebook? In any case, I asked it to get KaTeX working and it did- probably by fixing whatever config mistake I made previously. Overall this took ~20m of my time (spread out over maybe 2h where I was multi-tasking). Again, this felt about right. ## Task 4: `kern_gp` In 2024 I wrote a tiny python package for GP inference where the _kernel matrix_ is the input- allowing you to completely handle the logic of evaluating the kernel yourself. I called the package `kern_gp` ([link](https://github.com/AustinT/kernel-only-GP/)), and it was used for the experiments in the paper [Hash Collisions in Molecular Fingerprints: Effects on Property Prediction and Bayesian Optimization](https://arxiv.org/abs/2511.17078) which I co-authored last year with a student, Walter Virany. The package was left as a stub on my GitHub. I asked Claude to fix up the README, convert the package manager to uv, and write a GitHub action to publish the package to PyPI, which it did. The package is now live here: Overall this took ~10m only, which felt right. ## Conclusion Overall it felt really nice to tick a lot of these low-urgency tasks off of my to-do list. My only complaint is that I can't really multitask between supervising Claude and doing deep work. The best workflow I have is multitasking to do several of these tasks at once.