One of the interesting challenges of the AI ecosystem in 2026 is that new, effective patterns emerge faster than I can adopt them. I’ll find a handful, get back to work, and realize a month later that I’d missed four or five more. The adoption cycle for Imprint this year has been something like: January: get every engineer onto Claude Code every single day March: ok, let’s also get everyone else onto Claude Code or Claude Cowork every single day April: local development is bottlenecked on checkout and worktree model, instead create ~10 local workspaces which each have an independent checkout of every repository, and operate at the workspace level, not at the repository level, so it can generate cross-repository pull requests across frontend, backend, infrastructure and data monorepos June: oh boy, agent-driven development is heavily constrained by lack of a common task management system with higher visibility and less permission complexity than Jira, so let’s migrate the entire company over to Linear and hard stop on Jira July: yikes, now we have visibility into all these tickets, many of them are trivial but managing them through local development isn’t scaling, let’s roll out an orchestrated harness which internally we call “Agent Fleet”, along the lines of Stripe’s Minions The most recent question for me has been figuring out how to adopt the software factory pattern. (After some light research, the specific AI-context origin of this term is slightly messy to attribute, but I think it might be Justin McCarthy in February 2026’s Software Factories And The Agentic Moment.) The software factory pattern is looping on a broad goal, and then relying on the harness to drive progress towards that goal. Our first pass at implementation is fairly basic: An agent skill /linear-project-loop which reads in a Linear project and starts by auditing that project’s goal definition on these dimensions: An RFC in Notion that describes the project’s goals, how those goals are measured, and the general approach A Datadog dashboard or Snowflake queries that measure progress against those goals If those are missing, or the Linear project is missing in its entirety, it iterates with you on creating those missing tools. Then it reviews the state of the metrics and issues for the project. If new work is identified, it adds those issues to the project. It updates the state of issues that have moved. It works on the non-blocked tasks based on the project’s current state. This is often writing a pull request, updating a pull request, pinging for review, asking a clarifying question, etc. When a task completes, if the project description is fresh, it takes on the next task. If the description hasn’t been updated in a while, it reruns the loop starting with the first step. Right now I am running this locally in a local harness, but it’s working well enough that I anticipate moving the behavior to be driven by the same orchestrated harness that we assign one-off tasks to. What I particularly like about the factory pattern is that it parallels very closely how I’ve been working locally, while forcing me to recognize the places where I was accidentally hording parts of the state for myself regarding the goals of the project. I was already asking agents to iterate on specific Linear projects, but they didn’t have the ability to evaluate if they were going in the right direction, or if it was missing necessary tasks. Now it does. The other place this has been extremely helpful for me is checking in on projects post release. For example, I shipped our passkeys implementation earlier this year, but some months go by without my checking in on how it’s going. If we saw adoption spike, or error rates start to turn, I might miss it, but running the factory in a less frequent post-release mode would catch it immediately. The final thought that’s been interesting to me is how much all of the pieces here compound only to the extent that you have the other pieces. For example, this factory pattern depends on having Datadog MCP and Snowflake access available to manage goal-tracking, but it also depends on Linear being the single source of state for the company’s work, and an orchestrated harness that can perform work independently from your laptop. Keeping up with this many migrations is a fascinating industry moment.

I owe a lot of my professional identity and success to CSS-Tricks. CSS-Tricks repeatedly gave me the opportunity to write for them. In doing so, they helped to both socialize and normalize accessibility as a mainstream frontend concern. I’m deeply thankful to them for this. The team was also a joy to work with, notably Geoff Graham. He’s a mensch, and one of the nicest people you can interact with in the frontend web space. If you have not been following the news about the site, Kevin Powell has a good video about the whole situation: Content skipped. I’m not speaking on behalf of Geoff, Chris, or others involved with running the current version of CSS-Tricks. I’ve got skin in the game as an author. This is my personal opinion, born of my feelings and beliefs. I think a lot of the web’s infrastructure should be co-ops, and CSS-Tricks is knowledge infrastructure. To that point, I should also point out that the website covers far more than just CSS. The corporate model of ownership can be a risk. If infrastructure is not part of a corporation’s core strategy, it is not a priority. As Kevin’s video touched on, it seems like promotion via owning the frontend content space isn’t part of Digital Ocean’s strategy anymore. It is not that CSS-Tricks does not have value. It is that Digital Ocean cannot see it. It is deeply, tragically ironic to me that Digital Ocean allowed this to transpire. This is because I know for a fact that the techniques and philosophies shared by CSS-Trick authors helped to shape iterations of their product’s UI. Some may be quick to point out that this knowledge now—illegally—exists inside of LLM training data, so the risk of the website going away is mitigated. To this, know that we should be striving to keep resources like CSS-Tricks going. Human creativity is the force that creates new techniques, strategies, and technologies. The web will calcify without voices sharing what they know, forever locking us into endless permutations of a fixed point in time. Unlike corporations, co-ops don’t have to be motivated by profit. By not needing to prioritize growth at all costs it means co-ops can instead prioritize and incentivise things like preservation and cultivation. It is also a successful model of operation, one that even already exists, and flourishes in the tech space. Collective ownership can also serve as checks and balances for, and protection against hierarchical decision-making. I only need to point to the chaotic and aberrant decisions many CEOs in the technology space have been making as of late to demonstrate the value of this approach. Paddy Srinivasan, if you somehow wind up reading this: Save some face and take a big swing. Give CSS-Tricks back to the people who love it.

How can something that “just works” be so annoying? situation We live in Cambridge off a little road down a drive in shared ownership between us and our neighbouring houses. All the utilities are buried under this drive, including the phone line. anticipation Over the last few years we have been canvassed repeatedly by CityFibre saying that they can deliver fibre all way to our house. I saw them digging trenches and leaving tails of purple fibre cladding along nearby roads, ready to hook up all the houses. I thought they would need to do something similar to deliver fibre to us. So when they turned up and knocked on our door, I talked to their salesbods and walked them up and down the drive and pointed out where the existing BT line goes. Then they gave up trying to sell to us. This happened about three times. disaffection We were not eager enough for an upgrade to deal with these impediments. notification A few months ago we were told that CityFibre would soon come and do the upgrade, since there’s a nationwide deadline for turning off the copper phone network at the end of the year. We expected that this would force them to actually plan some digging works, so we talked to our neighbours about it. We were all ready for some huge faff to follow the next visit by the CityFibre bods. installation CityFibre turned up on the promised morning bright and early. To our enormous surprise, a brown fibre housing was already poking out of the ground next to our copper phone line. It had been fed through 50 metres of 5cm duct without us being aware they were even working on the street. Within a couple of hours, the technicians had drilled through our wall, installed the ONT, blown fibre through the unexpected pipe, plugged in the CPE (superficially identical to the old one), and left telling us to anticipate that it might not work properly until tomorrow. activation Around lunch time, the copper phone line stopped working completely. Some faff ensued, switching all our devices over to the new WiFi network. For a while we thought this was the death of our land line, but in the course of debugging other issues, I realised that the router has a built-in VoIP adapter (I don’t think we were told it has a built-in VoIP adapter) so I plugged the phone in and it Just Worked: they had ported our phone number across and everything. Flawless. I was seriously impressed. rumination It has been a few weeks since the switchover, and apart from a couple of horrible Clown-afflicted IoT devices, it has been fairly smooth. What prompted me to write this up was realising that we delayed this upgrade for years because the sales people were not given enough technical information about how the installation process works: the fact that houses typically have a 5cm duct containing the copper lines (probably standard for the last 40 years) and the fact that fibre can be shoved through a few tens of metres without difficulty. And worse, the sales people didn’t have an esclation path for difficult cases: they just gave up instead. From a technical point of view, the installation was impeccable. (I guess the loose 24 hour window for the cutover time was because OpenReach and CityFibre don’t have tight requirements on ISP reconfiguration schedules.) From the sales point of view, it was crap. Maybe it would have gone faster if we offered to switch early without asking if the drive would be a problem? But I guess the difference between “yes!” and “yes, but will this be a problem?” is too much to expect from a minimum-wage door-to-door salesbod whose employer didn’t give them enough information or any escalation path.

Read the post here.

I listen to a lot of podcasts, and I like how they fit around other tasks. I press play, lock my phone, and put it down. I’m free to wash the dishes, fold the laundry, or shop for groceries. Unfortunately, more and more information is only published as a video. Technical talks, conference sessions, video essays – they don’t work in an audio-only podcast app. I could convert these videos to MP3 files, but that breaks down the moment a video isn’t pure spoken word. If a speaker says, “Look at this slide” or holds up a diagram, an audio-only file leaves me stranded. I don’t want to give up the podcast player I like, nor stare at a screen for an hour – but I do want the information in these videos. To solve this, I’m abusing my podcast player’s chapter support. This gives me the best of both worlds: I can listen to a video as audio-first, and glance at my lock screen if I need a moment of visual context. The idea: Chapters every few seconds MP3 files can have ID3 metadata, and ID3 metadata can include chapters. A chapter covers a particular time range, and it can have an associated title, description, and cover art. My podcast app of choice is Overcast, which can’t play videos, but it does have robust chapter support. I can jump between chapters, navigate a table of contents, and see per-chapter cover art. To get videos into Overcast, I’m creating MP3 files with a new chapter every few seconds, and the per-chapter cover art is a corresponding frame from the video. As I play the file, I get a slow, stop-motion-like rendition of the original video. If my phone is locked, I can glance at my lock screen and see the current frame in the Now Playing screen. Overcast is developed by Marco Arment, and I got this idea from Forecast, his app for adding chapters to podcasts. In particular, I was struck by its ability to create chapters that don’t display in the chapter list – ideal if I don’t want a table of contents with hundreds of entries. As I was developing my script, I compared my output to the output from Forecast to ensure I was creating the chapters correctly. The code: FFmpeg and Mutagen There are three steps in this process: Convert a video file to an MP3 Extract images from the video at a fixed interval Insert the images as hidden chapters in the MP3 file Let’s go through each in turn. 1. Convert a video file to an MP3 Converting a video file to an MP3 is a single FFmpeg command: ffmpeg -i video.mp4 audio.mp3 This is consistently the slowest step of the process, and I do wonder if I could use different settings or an alternative encoder to make it go faster – but it’s not slow enough to be worth further investigation. 2. Extract images from the video at a fixed interval Extracting images from a video needs a more complicated FFmpeg command: ffmpeg -i video.mp4 \ -vf 'fps=1/5,scale=iw*sar:ih,scale=min(iw\,945):min(ih\,945):force_original_aspect_ratio=decrease' \ thumbnail_%04d.jpg This extracts an image every 5 seconds, downscales any image larger than 945 pixels square (while preserving the original aspect ratio), and saves the results as sequentially numbered JPEG images (thumbnail_0001.png, thumbnail_0002.png, and so on). The key is the -vf flag, which defines two FFmpeg filters: The fps filter selects one frame every 5 seconds (fps=1/5). The first scale filter scales the width based on the sample aspect ratio (scale=iw*sar:ih). Without this filter, frames can be stretched and distorted. The second scale filter scales the input video, preserving the original aspect ratio (force_original_aspect_ratio=decrease), and ensuring the output images fit within 945×945px or the size of the input video, whichever is smaller. My limit is 945 pixels because that’s the largest size that cover art is shown on my iPhone. This filter still isn’t completely correct – it sometimes creates images from portrait videos that are smaller than I’m expecting – but it’s good enough. These are only thumbnails for glancing at, and if I want to change it later, I can always do the image resizing outside FFmpeg. 3. Insert the images as hidden chapters in the MP3 file Inserting the chapters into the MP3 file is more complicated. Although FFmpeg has basic support for ID3 metadata, as far as I know, it can’t insert chapters with per-chapter artwork. Instead, I’m going to reach for Python and the Mutagen library. Here’s the code to add a chapter to an MP3 file: from mutagen.id3 import APIC, CHAP, ID3, PictureType audio = ID3("audio.mp3") with open("thumbnail_0001.jpg", "rb") as f: img_data = f.read() image_frame = APIC(mime="image/jpeg", type=PictureType.OTHER, data=img_data) chapter_frame = CHAP( element_id="chp1", start_time=0, end_time=5 * 1000, sub_frames=[image_frame] ) audio.add(chapter_frame) audio.save() This creates a single chapter that lasts the first 5 seconds (0 to 5000 milliseconds), and the per-chapter cover art is thumbnail_0001.jpg. If we ran this in a loop, we could add images for every 5 second slice of the original video. This code is inserting two frames into the ID3 metadata: The CHAP (chapter) frame contains the timing information, and it can have subframes for metadata like title, chapter art, or associated URL. The APIC (attached picture) subframe contains information about a picture, which can either be a blob of image data or a URL to an image on the web. Normally, you’d also insert a CTOC frame which defines a table of contents, but I don’t want a TOC with hundreds of 5-second chapters, so I’m deliberately not doing this here. This is allowed by the ID3 spec – you’re not required to insert a CTOC frame if you’re using chapters, and you can have chapters that aren’t listed in your table of contents. To work out which frames I needed, I used Forecast to create some chapters by hand, and I inspected their frames. In particular, loading an MP3 and calling Mutagen’s pprint() method shows a human-readable list of frames, and then I could drill into the individual fields: from mutagen.id3 import ID3 audio = ID3("audio.mp3") print(audio.pprint()) I wrapped all this code in a project called glancecast, which allows you to convert a video file with a single command, with optional flags to set the frame length and chapter art size: $ python3 glancecast.py interesting_talk.mp4 interesting_talk.mp3 The process takes a minute or so to complete, most of which is spent transcoding the video file to MP3. The resulting MP3s are usually 40 to 50 MB in size, which is very reasonable. The outcome: How it looks in practice Here’s what one of these “glanceable” podcasts looks like in Overcast and on my lock screen: Maggie Appleton presented this talk over two years ago and it’s been on my “talks to watch” list ever since. Once I put it in Overcast? I listened to it in less than a day. It’s not a lot of extra information, but enough that I can quickly glance down and get the gist of what a speaker is saying. Both views update with a new frame every few seconds, or I can put my phone in my pocket and ignore the screen. I’ve used this approach for half a dozen videos so far, and I’m happy with the results. I expect to keep using it, because I have a long queue of videos I’ve been meaning to watch. If you’d like to try this, check out glancecast for the full code and instructions. [If the formatting of this post looks odd in your feed reader, visit the original article]