Uses This

UsesThis.com has >1000 interviews answering the same 4 questions related to how people work.

Skip to the plots, starting at #OS Usage

Outside Inspiration

I'm trying to find setups that are outside. There are lots of matches, especially in dream setups. Some other useful search terms: sun, grass, porch

grep -P '\b(sun|weather|outside|porch)\b'  usesthis_pages/* |grep -Pv "outside (of|that|your)" | pandoc -f html -t plain
grep -P '\b(grass)'  usesthis_pages/* | grep -v Grasshopper | pandoc -f html -t plain

Ultimately, I didn't get an additional insight for my own setup. But did confirm many like working outside.

I also wish I'd symlinked the filenames to include date and the author categories. Then grep could have had some useful metatdata. As it is, I just jumped to the matching file.

More details

Now that there's 5Mb of interviews locally, I want to poke at it a bit.

Extracting

UsesThis links tools and hardware (with useful title attributes too). They can be pulled out like

<a[^>]+(?<title>title="[^"]+)[^>]*>(?<name>.*?)<\/a

But there are a few links on every page. I try to exclude a those

href="(/"|.*ko-fi.com|/about|/categories/")|Attribution-Share Alike|subscribing to the feed

Tool links occur anywhere in the 4 Questions asked:

Who are you, and what do you do?
What would be your dream setup?
What hardware do you use (first 14 are What do you use to get the job done?)
And what software?

I pull these into $sec (section; would be better named as question)

\QWho are you, and what do you do?\E|
\QWhat would be your dream setup?\E |
\QWhat hardware do you use\E        |
\QAnd what software?\E              |
\QWhat do you use to get the job done?\E

Publish time is also well annotated.

time datetime="([^"]+)

Extracting all in one pass with perl. This started as perl -lne 'BEGIN{} ...' usesthis_pages/* > usesthis_tags.txt, but to explore org mode/literate programming. I moved into noweb and perl.

open STDOUT, '>', "usesthis_tagsx.txt"; # bash '> usethis_tags.txt'
@ARGV = glob('usesthis_pages/*');       # 'perl -n' part 1
sub resetf {$dt=0; $sec="top";}         # first section and time
resetf;
while(<>){                              # 'perl -n' part 2
  $dt=$1 if m/<<regex-date>>/;
  $sec=$& if m/<<regex-sec>>/x;
  while(/<<regex-tag>>/g){
   $t=$+{title};
   $n=$+{name};
   next if $& =~ m:<<regex-exclude>>:;
   print("$dt\t",
         $ARGV=~s:.*/::r ,
         "\t$sec\t$n\t",
         $t=~s/.*?"|\.$//gr, "\n");
  }
  resetf if(eof);
}

Org-mode org-babel noweb aside

Originally, the perl-to-long-tag extraction was iterated on in a terminal using an expansive perl one-liner. Once it was running as expected, I run on the full glob (all files). Then removed the bash wildcard with perl glob() and > redirect with open STDOUT I also moved the regular expressions into their own code blocks and weaved them in with noweb (eg. <<regex-date>>).

But arrant angles << >> messed with perl syntax highlighting. I moved the pattern delimiters and modifiers m:...:x back into the code. This means that there's context need to understand the regexp that's only in the code:

x modifier for ignore spacing
m:...: colon deliminator so slash doesn't have to be escaped

What's more :noweb-prefix no is needed in the babel headers to correctly use the multi-line pattern.

Tool use

pacman::p_load(dplyr,readr,stringr, ggplot2)
clps <- \(x) x |> str_trunc(width=20) |> sort() |> unique() |>  paste0(collapse=";")

long <- read_delim("usesthis_tags.txt",
                   col_names=c("published", "who","sec","name","desc"),
                   delim="\t") |>
  mutate(published=lubridate::ymd(published)) |>
  filter(!is.na(published)) # 56 for whom name was before datetime

tags <- long |>
  group_by(name)|>
  summarise(n=n(), fdate=min(published), ldate=max(published),
            desc=clps(desc)) |>
  arrange(-n)

tags|>head()

name	n	fdate	ldate	desc
mac	948	2009-01-02	2026-09-26	All the mac inter…
developer	491	2009-01-02	2026-09-26	All the developer…;Charlie's Git…;Mitchell's Gi…
MacBook Pro	436	2009-01-02	2026-08-13	A laptop
Photoshop	378	2009-01-09	2026-04-30	A bitmap image ed…
windows	309	2009-01-17	2026-09-05	All the windows i…
Chrome	266	2009-01-17	2025-05-30	A web browser;A web browser app;A WebKit-based br…

OS

"top" section (before questions) contains category tags. These include operating system and biographical info ("developer", "artist"). Interviewees can and do endorse more than one of either.

os <- long |>
  filter(name %in% c('linux','mac','windows')) |>
  group_by(who) |>
  summarise(published=first(published),
	    os_long=clps(name),
            os=ifelse(grepl(';',os_long), "mutli", first(name)))

os |> count(os_long) |> arrange(-n)

os_long	n
mac	763
windows	166
mac;windows	93
linux	64
linux;mac	64
linux;mac;windows	28
linux;windows	22

To simplify, each interviewee is assigned the category they endorsed that is most often endorsed by others: artist, developer becomes just developer.

cats <- long |>
  filter(sec=="top", ! name %in% c('mac','linux','windows')) |>
  count(name)

main_desc <- cats |> inner_join(long) |>
  group_by(who) |> filter(n==max(n)) |> ungroup()

category_cnt <- main_desc |> count(name) |> arrange(-n)
head(category_cnt)

name	n
developer	491
writer	212
designer	158
artist	76
musician	60
game	42

Who uses linux?

cat_os <- main_desc |> inner_join(os)

# user of linux (alone or w/mac and or win)
# 'name' is catagory (dev, lawyer, hacker, etc)
cat_os |> select(-n) |>
  filter(grepl('linux',os_long)) |>
  count(name, name='linux') |>
  # merge to catagory counts (across all endoursments, many per person)
  inner_join(cats)  |>
  # calc percent
  transmute(name,linux, total=n, percent=round(linux/total*100,2)) |>
  arrange(-percent) |>
  head(n=10)

name	linux	total	percent
developer	139	489	28.43
service	2	8	25
hacker	8	33	24.24
lawyer	1	5	20
politics	1	5	20
professor	5	33	15.15
scientist	5	33	15.15
engineer	1	11	9.09
hardware	1	11	9.09
researcher	3	35	8.57

OS Usage

p_os <- long |>
  filter(name %in% c('linux','mac','windows')) |>
ggplot() + aes(x=published, fill=name) +
  geom_histogram(position="stack",bins=10) +
  theme_bw() + facet_wrap(~name) + labs(title="OS by interview", fill='os')

p_cat <- cat_os |>
  filter(name %in% category_cnt$name[1:5]) |>
 ggplot() + aes(x=published, fill=os) + geom_histogram() +
 theme_bw() + facet_wrap(~name)

pacman::p_load(patchwork)
(p_os+guides(fill="none")+labs(x="")) / p_cat +
  plot_layout(guides = "collect", axis="collect", heights = c(1, 2))

../images/usesthis/os.png

Software

I'm glad firefox isn't lagging too far behind chrome. And emacs and vim are top editors.

tools <- long |>
  filter(grepl('what software|get the job', sec)) |>
  transmute(who,
            tool=tolower(name) |> gsub('[^a-z0-9]','',x=_)|>
	      gsub('googlechrome','chrome',x=_)|>
	      gsub('mozillafirefox','firefox',x=_)|>
	      gsub('sublimet.*','sublime',x=_)|>
	      gsub('gnuemacs','emacs',x=_) |>
	      gsub('i3tile.*','i3',x=_) |>
	      gsub('rhymeswithgonad','xmonad',x=_) |>
	      gsub('visualstudiocode','vscode',x=_) |>
	      gsub('intellijidea','intellij',x=_) |>
	      gsub('^vi$','vim',x=_) |>
	      gsub('awesomew.*','awesome',x=_),
	    desc_cat=desc|> tolower() |>
	      gsub('irc ','chat',x=_)|>
	      gsub('ide\\b','text editor ',x=_)|>
	      gsub('terminal shell','shell',x=_)|>
              str_extract('chat|browser|window manager|text editor|term|shelll')) |>
  merge(cat_os, by='who')

linux_tools <- tools |>
  filter(grepl('linux',os_long)) |>
  count(tool, name, desc_cat) |>
  arrange(-n)

tools_totalsum <- linux_tools |>
  group_by(tool) |>
  summarise(nsum=sum(n)) |>
  merge(linux_tools, by="tool") |>
  filter(!is.na(desc_cat),nsum>2)

ggplot(tools_totalsum) +
 aes(x=tool, fill=name, y=n) +
  geom_histogram(stat='identity') +
  facet_wrap(~desc_cat, scales="free_x") +
  see::theme_modern(axis.text.angle=45)

../images/usesthis/tools.png

Editors

Over time we can see vscode take off, emacs ebb and flow, atom pop in and out, eclipse fade away.

all_tools <- tools |> filter(desc_cat %in% c('text editor'))

all_tools |> count(tool,name='ntool') |>
  merge(all_tools, by='tool') |>
  filter(tool %in% tools_totalsum$tool) |>
  #filter(ntool>2) |>
ggplot() + aes(x=published, y=tool, color=os) + geom_point() +
facet_grid(desc_cat~., scales="free") + theme_bw() +
labs(title="Text editors over time")

../images/usesthis/editor.png

LLM inference usage

prompted chatgpt for skip-existing option in curl --remote-name-all, circled around --no-clobber to decided that wouldn't work (makes *.1 instead of overwrite)
perl -lne '...' gl*ob > out to @ARGV=...; open STDOUT, ... was a mix of chatGPT hallucination, Claude, stack overflow, and perldoc.
uploaded pandoc converted plaintext file of all to query with sonnet 5 medium. prompt for summary, specific hardware used outside. confirmed what was found with grep.