👨‍💻 Wesley Moore

Turning One Hundred Tweets Into a Blog Post


Near the conclusion of my #100binaries Twitter series I started working on the blog post that contained all the tweets. It ended up posing a number of interesting challenges and design decisions, as well as a couple of Rust binaries. Whilst I don’t think the process was optimal I thought I’d share the process to show my approach to solving the problem. Perhaps the tools used and approach taken is interesting to others.

My initial plan was to use Twitter embeds. Given a tweet URL it’s fairly easy to turn it into some HTML markup. By including Twitter’s embed JavaScript on the page the markup turns into rich Twitter embed. However there were a few things I didn’t like about this option:

So I decided I’d render the content myself. I also decided that I’d host the original screenshots and videos instead of saving them from the tweets. This was a little time consuming as they were across a couple of computers and not named well but I found them all in the end.

To ensure the page wasn’t enormous I used the loading="lazy" attribute on images. This is a relatively new attribute that tells the browser to delay loading of images until they’re within some threshold of the view port. It currently works in Firefox and Chrome.

I used preload="none" on videos to ensure video data was only loaded if the visitor attempted to play it.

To prevent the blog post from being too long/heavy I split it across two pages.

Collecting All the Tweet URLs

With the plan in mind the first step was getting the full list of tweets. For better or worse I decided to avoid using any of Twitter’s APIs that require authentication. Instead I turned to nitter (an alternative Twitter front-end) for its simple markup and JS free rendering.

For each page of search results for ‘#100binaries from:@wezm’ I ran the following in the JS Console in Firefox:

tweets = []
document.querySelectorAll('.tweet-date a').forEach(a => tweets.push(a.href))

and pasted the result into tweets.txt in Neovim.

When all pages had be processed I turned the nitter.net URLs back in to twitter.com URLs: :%s/nitter\.net/twitter.com/.

This tells Neovim: for every line (%) substitute (s) nitter.net with twitter.com.

Turning Tweet URLs Into Tweet Content

Now I needed to turn the tweet URLs into tweet content. In hindsight it may have been better to use Twitter’s GET statuses/show/:id API to do this (possibly via twurl) but that is not what I did. Onwards!

I used the unauthenticated oEmbed API to get some markup for each tweet. xargs was used to take a line from tweets.txt and make the API (HTTP) request with curl:

xargs -I '{url}' -a tweets.txt -n 1 curl https://api.twitter.com/1/statuses/oembed.json\?omit_script\=true\&dnt\=true\&lang\=en\&url\=\{url\} > tweets.json

This tells xargs to replace occurrences of {url} in the command with a line (-n 1) read from tweets.txt (-a tweets.txt).

The result of one of these API requests is JSON like this (formatted with jq for readability):

  "url": "https://twitter.com/wezm/status/1322855912076386304",
  "author_name": "Wesley Moore",
  "author_url": "https://twitter.com/wezm",
  "html": "<blockquote class=\"twitter-tweet\" data-lang=\"en\" data-dnt=\"true\"><p lang=\"en\" dir=\"ltr\">Day 100 of <a href=\"https://twitter.com/hashtag/100binaries?src=hash&amp;ref_src=twsrc%5Etfw\">#100binaries</a><br><br>Today I&#39;m featuring the Rust compiler — the binary that made the previous 99 fast, efficient, user-friendly, easy-to-build, and reliable binaries possible.<br><br>Thanks to all the people that have worked on it past, present, and future. <a href=\"https://t.co/aBEdLE87eq\">https://t.co/aBEdLE87eq</a> <a href=\"https://t.co/jzyJtIMGn1\">pic.twitter.com/jzyJtIMGn1</a></p>&mdash; Wesley Moore (@wezm) <a href=\"https://twitter.com/wezm/status/1322855912076386304?ref_src=twsrc%5Etfw\">November 1, 2020</a></blockquote>\n",
  "width": 550,
  "height": null,
  "type": "rich",
  "cache_age": "3153600000",
  "provider_name": "Twitter",
  "provider_url": "https://twitter.com",
  "version": "1.0"

The output from xargs is lots of these JSON objects all concatenated together. I needed to turn tweets.json into an array of objects to make it valid JSON. I opened up the file in Neovim and:

I then reversed the order of the objects and formatted the document with jq (from within Neovim): %!jq '.|reverse' -.

This filters the whole file though a command (%!). The command is jq and it filters the entire document ., read from stdin (-), through the reverse filter to reverse the order of the array. jq automatically pretty prints.

It would have been better to have reversed tweets.txt but I didn’t realise they were in reverse chronological ordering until this point and doing it this way avoided making another 100 HTTP requests.

Rendering tweets.json

I created a custom Zola shortcode, tweet_list that reads tweets.json and renders each item in an ordered list. It evolved over time as I kept adding more information to the JSON file. It allowed me to see how the blog post looked as I implemented the following improvements.

The markup Twitter returns is full of t.co redirect links. I wanted to avoid sending my visitors through the Twitter redirect so I needed to expand these links to their target. I whipped up a little Rust program to do this: expand-t-co. It finds all t.co links with a regex (https://t\.co/[a-zA-Z0-9]+) and replaces each occurrence with the target of the link.

The target URL is determined by making making a HTTP HEAD request for the t.co URL and noting the value of the Location header. The tool caches the result in a HashMap to avoid repeating a request for the same t.co URL if it’s encountered again.

I used the ureq crate to make the HTTP requests. Arguably it would have been better to use an async client so that more requests were made in parallel but that was added complexity I didn’t want to deal with for a mostly one-off program.

Adding the Media

At this point I did a lot of manual work to find all the screenshots and videos that I shared in the tweets and added them to my blog. I also renamed them after the tool they depicted. As part of this process I noted the source of media files that I didn’t create in a "media_source" key in tweets.json so that I could attribute them. I also added a "media" key with the name of the media file for each binary.

Some of the externally sourced images were animated GIFs, which lack playback controls and are very inefficient file size wise. Whenever I encountered an animated GIF I converted it to an MP4 with ffmpeg, resulting in large space savings:

ffmpeg -i ~/Downloads/so.gif -movflags faststart -pix_fmt yuv420p -vf "scale=trunc(iw/2)*2:trunc(ih/2)*2" so.mp4

This converts so.gif to so.mp4 and ensures the dimensions are a divisible by 2, which is apparently a requirement of H.264 streams encapsulated in MP4. I worked out how to do this from: https://unix.stackexchange.com/a/294892/5444

I also wanted to know the media dimensions for each file so that I could have them scaled properly on the page — most images are HiDPI and need to be presented at half their pixel width to appear the right size.

For this I used ffprobe, which is part of ffmpeg. I originally planned to use another tool to handle images (as opposed to videos) but it turns out ffprobe handles them too.

Since I wanted to update the values of JSON objects in tweets.json I opted to parse the JSON this time. Again I whipped up a little Rust “script”: add-media-dimensions. It parses tweets.json and for each object in the array runs ffprobe on the media file, like this:

ffprobe -v quiet -print_format json -show_format -show_streams file.mp4

I learned how to do this from: https://stackoverflow.com/a/11236144/38820

With this invocation ffprobe produces JSON so add-media-dimensions also parses that and adds the width and height values to tweets.json. At the end the updated JSON document is printed to stdout. This turned out to be a handy sanity check as it detected a couple of copy/paste errors and typos in the manually added "media" values.

The oEmbed markup that Twitter returns includes links for each piece of media. Now that I’m handling that myself these can be deleted. Neovim is used for this:

:%s/ <a href=\\"https:\/\/twitter\.com[^"]\+\(photo\|video\)[^"]\+">pic.twitter.com[^<]\+<\/a>//

For each line of the file (%) substitute (s) matches with nothing. And that took care of them. Yes I’m matching HTML with a regex, no you shouldn’t do this for something that’s part of a program. For one-off text editing it’s fine though, especially since you can eyeball the differences with git diff, or in my case tig status.

Adding a HiDPI Flag

I initially tried using a heuristic in tweet_list to determine if a media file was HiDPI or not but there were a few exceptions to the rule. I decided to add a "hidpi" value to the JSON to indicate if it was HiDPI media or not. A bit of trial and error with jq led to this:

jq 'map(. + if .width > 776 then {hidpi: true} else {hidpi:false} end)' tweets.json > tweets-hidpi.json

If the image is greater then 776 pixels wide then set the hidpi property to true, otherwise false. 776 was picked via visual inspection of the rendered page. Once satisfied with the result I examined the rendered result and flipped the hidpi value on some items where the heuristic was wrong.

Adding alt Text

Di, ever my good conscience when it comes to such things enquired at one point if I’d added alt text to the images. I was on the fence since the images were mostly there to show what the tools looked like — I didn’t think they were really essential content — but she made a good argument for including some alt text even if it was fairly simplistic.

I turned to jq again to add a basic "media_description" to the JSON, which tweet_list would include as alt text:

jq 'map(. + {media_description: ("Screenshot of " + (.media // "????" | sub(".(png|gif|mp4|jpg)$"; "")) + " running in a terminal.")})' tweets.json > tweets-alt.json

For each object in the JSON array it adds a media_description key with a value derived from the media key (the file name with the extension removed). If the object doesn’t have a media value then it is defaulted to “????” (.media // "????").

After these initial descriptions were added I went though the rendered page and updated the text of items where the description was incorrect or inadequate.

Video Poster Images

As it stood all the videos were just white boxes with playback controls since I has used preload="none" to limit the data usage of the page. I decided to pay the cost of the larger page weight and add poster images to each of the videos. I used ffmpeg to extract the first frame of each video as a PNG:

for m in *.mp4; do ffmpeg -i $m -vf "select=1" -vframes 1 $m.png; done

I learned how to do this from: https://superuser.com/a/1010108

I then converted the PNGs to JPEGs for smaller files. I could have generated JPEGs directly from ffmpeg but I didn’t know how to control the quality — I wanted a relatively low quality for smaller files.

for f in *.mp4.png; do convert "$f" -quality 60 $f.jpg ; done

This produced files named filename.mp4.png.jpg. I’m yet to memorise how to manipulate file extensions in zsh, despite having been told how to do it, so I did a follow up step to rename them:

for f in *.mp4; do mv $f.png.jpg $f.jpg ; done

Wrapping Up

Lastly I ran pngcrush on all of the PNGs. It reliably reduces the file size in a lossless manner:

for f in *.png; do pngcrush -reduce -ow $f; done

With that I did some styling tweaks, added a little commentary and published the page.

If you made it this far, thanks for sticking with it to the end. I’m not sure how interesting or useful this post is but if you liked it let me know and I might do more like it in the future.

Stay in touch!

Follow me on the Fediverse, subscribe to the feed, or send me an email.