CLAUDE.md. ~/.claude/ symlinked into ~/git/dotfiles/claude/)I use Ghostty for its simplicity, speed, and native macOS integration. I used to use tmux (see tmux + Neovim setup), but tmux adds another layer of latency and memory overhead. Now I just use Ghostty’s built-in splits: ⌘+D / ⌘+⇧+D.
My default layout is two panes side by side — Claude Code on the left, Neovim on the right. I split further as needed but try to return to this two-pane setup. Neovim is the right fit here for the same reason as Ghostty: snappy and versatile. You can quickly move around, explore files, and make edits while CC works in the other pane. I sometimes run watch scripts (e.g., auto-compile, file sync) in a tiny extra pane.
Usually I have several tabs open (one for each project) with each tab having 2+ panes.
Three layers:
- Global (~/.claude/CLAUDE.md): Python/uv preferences, wiki/blog paths, project tracking conventions, academic writing style
- Project (CLAUDE.md): repo-specific instructions, commands, code style
I keep a small set of research-oriented skills. They include:
math — symbolic verification with SymPydoi-bibtex — fetch BibTeX from DOIarxiv-metadata — fetch paper metadata from arXivopenalex-database — query OpenAlex for literature search and bibliometricsPlus the Anthropic document-skills plugin for PDF, PPTX, DOCX, etc.
My custom agents include:
file-paper — file an academic paper to my wiki. It fills in metadata and then search my wiki to identify & link relevant papers and other conceptswikify — turn anything into a wiki page. This calls file-paper agent when necessary. Otherwise follow similar steps to create links to relevant pages. latex-cleanup — review LaTeX documents after editingcodebase-analyzer, code-smell-refactor, security-auditor — code qualitygit-sync — add, commit, pull –rebase, pushrelease — bump version, tag, push, create GitHub release, publishuv-init — bootstrap Python/uv/direnv environmentnew-research-project — scaffold from my project templatewikify — routes to the right agent (file-paper for papers, wikify for everything else)See the Official documentation for more details.
Skills are stored as SKILL.md files in .claude/skills/SKILL_NAME/ (project-level) or ~/.claude/skills/SKILL_NAME/ (global). Each skill file contains a YAML header with a description of when the skill should be used, step-by-step instructions for CC to follow, and optional references to helper scripts.
The YAML header requires:
name: A short identifier for the skilldescription: When the skill should be invoked (CC uses this to decide relevance)When CC is launched, only the YAML headers of available skills are loaded to CC’s context. They are used as an index for CC. CC uses these descriptions to determine which skills to invoke.
~/.claude/skills/math/
├── SKILL.md
└── verify.py
The SKILL.md file may look like:
---
name: math
description: >
Verify mathematical derivations step-by-step using SymPy.
Use this when users ask to verify proofs, check calculations,
or solve symbolic math problems.
---
When the user asks you to verify or solve mathematical problems:
1. Use the bundled `verify.py` script for symbolic computation
2. Show step-by-step derivation
3. Verify the final result
## Usage
Run the verification script:
# Check two expressions are equal
~/.claude/skills/math/verify.py eq "EXPR1" "EXPR2"
# Simplify/expand expression
~/.claude/skills/math/verify.py simp "EXPR"
Always explain your reasoning alongside the output.
A skill may or may not have accompanying scripts. It can have no script at all (just instructions) or simple inline code examples.
Skills are not invoked by the user, but only by CC. Skills are exposed to CC based on the YAML header. When CC determines that the skill is applicable, it will invoke the skill. When invoked, CC reads the full skill file and follows its instructions.
The “math skill” shown above would be invoked automatically when you ask CC to check derivation of some mathematical formula.
Skills essentially let you encode repeatable workflows as reusable “recipes” that CC can execute on demand.
CC provides a built-in plugin marketplace for easy skill installation. Use the /plugin marketplace add command:
/plugin marketplace add <github-repo>
For example, to install the official Anthropic skills:
/plugin marketplace add anthropics/skills
This command:
github.com/anthropics/skills).md and associated scripts).claude/plugins/ directoryNote that these skills are not automatically available. You need to explicitly enable them by invoking /plugin command or manually editing .claude/settings.json (project-level) or ~/.claude/settings.json (global):
{
"enabledPlugins": {
"pyright-lsp@claude-plugins-official": true,
"document-skills@anthropic-agent-skills": true,
"example-skills@anthropic-agent-skills": true
}
}
In CC, you can check available skills by running:
/skills
If the skills are correctly installed, you should see the list of available skills. They will be invoked when necessary.
Several community repositories provide curated collections of ready-to-use skills:
Key features: native macOS integration (splits, tabs, fullscreen behave like a real Mac app), GPU rendering via Metal/OpenGL, fast text throughput (competitive with Alacritty), and modern features. Config is a single plain-text file.
I use it as my primary terminal — see Claude Code: How I use it.
]]>Most LaTeX CVs (pdflatex + article class) produce untagged PDFs with no structure metadata, which are invisible to screen readers. The fix is minimal – a few lines added to your preamble, switch to LuaLaTeX, and your existing layout is preserved.
Validated with PDFix PDF/UA validator.
\DocumentMetadata before \documentclass#This must be the very first thing in your .tex file, before \documentclass.
\DocumentMetadata{
lang = en,
pdfstandard = a-2b,
pdfstandard = ua-1,
pdfversion = 2.0,
testphase = {phase-II},
}
\documentclass[10pt]{article} % your existing documentclass
pdfstandard = a-2b – PDF/A-2b archival compliancepdfstandard = ua-1 – PDF/UA-1 universal accessibility compliancetestphase = {phase-II} – activates automatic tagging of headings, lists, and other structuresLuaLaTeX uses OpenType fonts via fontspec instead of the old Type1 font system.
% REMOVE these:
% \usepackage[utf8]{inputenc}
% \usepackage[T1]{fontenc}
% \usepackage[sc]{mathpazo}
% ADD these:
\usepackage{fontspec}
\setmainfont{TeX Gyre Pagella} % Palatino equivalent
\setmonofont{Inconsolatazi4-Regular.otf}[BoldFont=Inconsolatazi4-Bold.otf] % optional, for \texttt
Common font substitutions:
- mathpazo / Palatino -> TeX Gyre Pagella
- Times / mathptmx -> TeX Gyre Termes
- Computer Modern -> Latin Modern (already default in LuaLaTeX)
- Helvetica / helvet -> TeX Gyre Heros
\hypersetup#\hypersetup{
pdftitle={Curriculum Vitae - Your Name},
pdfauthor={Your Name},
pdfsubject={Academic CV},
pdflang={en},
pdfdisplaydoctitle=true, % required for PDF/UA
% ... your existing options (colorlinks, etc.)
}
pdfdisplaydoctitle=true is required by PDF/UA – it tells PDF viewers to show the document title in the title bar instead of the filename.
Packages like marvosym produce glyphs that can’t be mapped to Unicode, which fails PDF/UA validation. Use fontawesome5 instead:
% REMOVE:
% \usepackage{marvosym} % \Letter, \Mundus lack Unicode mapping
% ADD:
\usepackage{fontawesome5} % \faEnvelope, \faGlobe, \faOrcid, etc.
Icon examples:
- \faEnvelope – envelope (email)
- \faGlobe – globe (website)
- \faOrcid – ORCID logo
- \faGithub – GitHub logo
- \faPhone – phone
latexmk -lualatex cv
Or in a Makefile:
cv.pdf: cv.tex
latexmk -lualatex cv
pdfinfo cv.pdf – quick check that Tagged says “yes” (requires poppler: brew install poppler)phase-III tagging (full paragraph tagging) may conflict with some packages – phase-II is the safe choicepoppler for pdfinfo CLI verification (brew install poppler on macOS)\marginpar sidebar section layouts – works, layout preservedbiblatex with biber – worksfancyhdr – worksparalist (compactitem) – worksorcidlink – worksfontawesome5 – worksShift + Tab to cycle through normal, plan, and accept-edits mode.@ to mention specific files.# to add something to memory. Shift+Tab cycles through three modes: default, auto-accept edits, and plan. bypassPermissions is intentionally excluded from this cycle by default.
If you’re comfortable using bypass permissions mode regularly, set it as your default in ~/.claude/settings.json:
{
"permissions": {
"defaultMode": "bypassPermissions"
}
}
With this config, Claude Code starts in bypass permissions mode, and Shift+Tab will include it in the cycle so you can freely switch between modes.
tmux + Neovim Setup#If you prefer terminal-based light-weight setup over IDEs like cursor, a nice way is to have a tmux session per project with Claude Code and an editor (e.g., Neovim) open side by side. You read, check, and edit with Neovim while asking things to Claude Code. When doing this, it is critical for Neovim to auto-reload files when Claude Code edits them externally. Without this setup, you’ll be viewing stale files unless you force-reload with :e!, which creates an annoying situation with unsaved edits in the current buffer that diverged from the file content.
To do so, you have to configure tmux correctly in addition to Neovim.
Neovim configuration#Add to your Neovim config (e.g., lua/user/options.lua):
-- Auto-reload files when changed externally
vim.o.autoread = true
vim.api.nvim_create_autocmd({ 'FocusGained', 'BufEnter' }, {
command = 'checktime',
})
tmux configuration#Add to your tmux.conf:
set -g focus-events on
This tells tmux to forward focus events to Neovim. Without it, Neovim won’t know when you switch back to it and won’t trigger the FocusGained autocmd.
For Neovim, restart or source your config. For tmux, either restart tmux, or run tmux set -g focus-events on in the terminal, or run Ctrl+a (or your prefix) then : and type set -g focus-events on inside tmux.
I did a back-of-the-envelope calculation. Even though I thought I already fully appreciated the cost of car ownership, the result was eye-opening. Replacing my car with an e-bike was probably one of the most impactful and best financial decisions in my life.
Here’s the quick calculation. I take the average number: the average cost of owning a new car in the US is, according to AAA, roughly $12,000 per year or $1,000 per month1. If you have never thought about it, this may sound crazy. You may be confident that you are paying way less. But, note that car-related spending is usually the most underestimated category of spending. We tend to only pay attention to the fuel and not to other costs like insurance, registration, tax, and depreciation.
Why do we underestimate car costs so badly? First, people pay a big upfront cost that fades from memory. Depreciation is even harder to feel—your car quietly loses value every day, but you don’t see that depreciation. Second, we tend to unconsciously replace “driving cost” with fuel cost, forgetting everything else. Third, irregular expenses like repairs, registration, and taxes are infrequent enough to feel like surprises rather than actual, predictable costs of ownership. And then there’s parking; many of us pay for parking passes, meters, or garage fees so routinely that we stop even noticing it as a car cost. For instance, you pay somewhere between $400 and $1200 for parking garage access at UVA (vs. you can get $400+ incentive if you forgo the parking pass).
The result is that we massively underestimate how much we spend on cars.
tl;dr: ~2 cars x $10k/year x 5 years ~ $100k. Cost of e-bike ownership is a rounding error.
I bought my e-bike in 2020 and sold an old car. If I followed the American lifestyle, I could have purchased a new car instead. If we think about 5 years, I’d have spent about
$12,000x 5 =$60,000
to own a car. For an ebike, I’d use $100-$300/year. This includes depreciation, other equipment and accessories, maintenance, etc. You may be wondering why this is so low given that some ebikes are $2,000+ or even approaching $10,000. But the cost of owning an ebike is heavily concentrated in the purchase price. After buying one, you don’t really have much to spend on.
Although those expensive ebikes may be more visible, you can buy decent ebikes around $1,000 these days. In my case, I bought a Rad Runner Plus under $2,000 and sold it around half of that price. So, annualized depreciation was less than $200. Maintenance is just once or twice per year at less than $100 each time. The accessories and other equipment are also not too expensive; I still use many of the accessories that I bought. And the price of “fuel” is simply too cheap to be meaningful—maybe several dollars per year.
So, at the five-year mark, with $200/yr, the total cost for an ebike is really a rounding error at around $1,000. The difference would be
$60,000-$1,000≈$59,000
But that’s just one car. In the US, it’s totally normal for a household to have two or three. I could have also bought a car for my oldest as soon as they were old enough to drive.
$59,000+$12,000x 2 ≈$83,000
If we had replaced the other old car with a shiny new car, we could have spent maybe an additional $2,000 per year.
$83,000+$2,000x 5 ≈$93,000
That’s close to $100,000 right there just in 5 years!
$100,000 can be life-changing. It can mean being able to purchase a house or not—this was indeed a significant fraction of my recent down payment. It can mean a fully-paid college vs. a massive student debt. And if you invest the whole amount in an index fund (assuming 8%) for 30 years, it’ll grow to ~$1M!
In other words, choosing whether or not to own a car can literally determine whether you’ll be a millionaire when you retire. And remember, this difference can be made in just five years.
What happens if we extrapolate? After putting the initial $100,000 to work, if you forego two cars (~$24,000/year) for 30 years, the difference grows to about ~$4M!
You might think: “But that’s for new cars. I drive a used car, so it’s way cheaper.” It’s true that the AAA number is based on new vehicles, and used cars do cost less, mainly because depreciation is lower. But maintenance and repair costs tend to be higher. When you add it all up, used car ownership still typically runs $5,000+ per year.
Also take a look at this video: Car ownership cost me a LOT of money. I’m done. He diligently collected most of the receipts and calculated how much he spent on his car. Even though he got his car almost for free, ended up spending over $4,000 a year. Most people buy much better (with plenty of remaining depreciation) used cars than this.
So let’s redo the math with used cars. Say, conservatively, you spend $6,000/year per car instead of $12,000. For two cars over 5 years, that’s still $60,000—and invested at 8% for 30 years, that alone grows to about ~$600,000. Similarly, starting with $60,000 and then saving $6,000 per year will get you to about $1.3M.
In other words, used cars can be definitely cheaper, but not drastically cheaper. It is still a million-dollar decision.
EVs are way better than gas cars, for the environment, for air quality, and for fuel and maintenance costs. But EVs still don’t drastically change the ownership cost picture. The upfront price and depreciation are still significant, and you still pay for insurance, registration, parking, and tires. Unless you’re lucky enough to lock in an insane lease deal or something, you’re still spending thousands of dollars a year to own one.
And from the perspective of cities and built environment, an EV is still a car. It still takes up the same amount of road and parking space, still contributes to congestion, still prevents you from moving, still makes streets less safe for pedestrians, and still fuels the same sprawling infrastructure. And it probably destroys the road more than gas cars. Switching from gas to electric is a step forward, but it doesn’t address lots of negative externalities of car ownership.
But even this is underestimating the difference. Cycling is one of the healthiest habits you can have—an excellent low-impact aerobic exercise. Some people worry about not getting exercise on an ebike, but studies show that e-bike riders get roughly the same amount of exercise as acoustic bike riders2. The reason is simple: because e-bikes are easier to ride, people ride them way more often!
It’s not just about physical health, but mental health too. From my experience, bike commuting can have enormous impact on mental health. The way it resets stress and lifts mood feels almost magical.
A report estimates that the health benefits of switching to bike commuting are about $1,000-$2,000 per year, although I’m not entirely sure whether this captures the full benefits of being healthy and happy.
And then there’s the impact on the safety of our streets—you’re not going to kill someone or destroy buildings with an ebike—and the climate crisis. These are harder to put a dollar figure on, but they’re enormous.
The benefits extend well beyond individual finances and fitness. If more people make this switch, many positive effects ripple through our cities.
Infrastructure costs can drop dramatically. Roads are expensive to build and maintain, and the damage grows exponentially3 with vehicle weight. A single truck does thousands of times more damage to the road than a bike. Better yet, bike lanes move five times more people per meter of street width than car lanes4. In other words, the number of people transported on a 5-lane stroad—that have been built and maintained with billions and billions of dollars—can be handled by a single two-way bike lane that costs almost nothing compared to that. Cities could redirect millions and billions from road maintenance to other priorities—or simply lower taxes.
Parking demand shrinks. Parking requirements have shaped American cities for decades, burying valuable urban land under asphalt. Constructing parking space is expensive. A single spot in a structured parking garage can cost $60,000 – $150,000 or even more for expensive places. Free parking seems free but someone always pays—usually through higher prices, higher taxes, or both. When people switch to bikes, that land becomes available for housing, parks, or businesses. And on a personal level, parking is a surprisingly concrete entry point: a university parking pass alone can run $500–$1,000 a year—enough to pay for a decent e-bike within a year or two.
Traffic congestion eases. Here’s the counterintuitive part: because bikes are so much more space-efficient than cars, it can trigger so-called traffic evaporation, where reduced road capacity actually leads to less total traffic. In Cambridge, MA, after separated bike lanes were installed on Garden Street, bike volumes surged over 500% with a corresponding drop in driving mode share. Across the Boston area, streets with new bike lanes saw car traffic drop 9–15% while bike trips soared. Every person switching from a car to other modes means less traffic because all other modes are much more space-efficient. If you have been to places like Copenhagen or Amsterdam, you must have seen all the quiet streets with very few cars (but with bike traffic jams). If some of those people on the bikes had to drive, these cities would have been completely choked with traffic.
The individual choice to ditch a car is already financially compelling. But when many people make that choice, the collective benefits multiply tremendously.
Here’s the bottom line: switching from a car-dependent average-American lifestyle to an e-bike probably saved my household up to $100k just over five years. If invested over a lifetime, that difference grows to over a million dollars.
Of course, this isn’t for everyone. Some people genuinely need cars for work, family situations, or disabilities. Some places just do not have any viable alternative to cars (~~run away!~~). I’m not arguing everyone should sell their car tomorrow. I’m just sad that our built environment, engineers, and car monoculture in North America pushes people into expensive car ownership.
But I think the most important step isn’t even about the numbers—it’s about questioning the default. Most of us never stop to ask: “Do I actually have to own a car?” We just assume we do, because everyone around us has one, because our cities were built around them, because it’s what you do when you grow up. But once you pause and genuinely ask the question, you might find that the answer is not as obvious as you thought.
I would argue that many of “you” are probably overestimating the difficulty of going car-light/car-free, while heavily underestimating the true cost of car ownership. Maybe you have never estimated the true cost, are too afraid to take the leap, or simply never imagined the alternatives.
So why not try your own back-of-the-envelope calculation? You might surprise yourself. And maybe try biking.
(share your thoughts on bluesky and subscribe 👇)
https://newsroom.aaa.com/wp-content/uploads/2025/09/UPDATE-AAA-Fact-Sheet-Your-Driving-Cost-9.2025-1.pdf ↩
Bourne et al. (2018). “Health benefits of electrically-assisted cycling.” International Journal of Behavioral Nutrition and Physical Activity. https://doi.org/10.1186/s12966-018-0751-8 ↩
Technically, road damage follows a power law (approximately the fourth power of axle load), not an exponential relationship. This was established by the AASHO Road Test (1958–1960). ↩
NACTO. “Transit Street Design Guide: Volume and Throughput.” https://nacto.org/publication/transit-street-design-guide/introduction/why/designing-move-people/ ↩
LLMs, Software engineering, Benchmark
Evaluated whether repository-level context files (e.g., AGENTS.md, CLAUDE.md) actually help coding agents. Found that they tend to reduce task success rates while increasing inference cost by over 20% — the files encouraged agents to explore more broadly and test more thoroughly without improving outcomes. The recommendation is to keep context files minimal.
Directly relevant to Claude Code’s use of CLAUDE.md. A cautionary example of how imposing human-written structure on capable agents can backfire. Over-specifying context constrains rather than assists (see Claude Code).