Scott Watermasysk - August 15, 2026
I wonder if “shared prompting” could help counter the deep code understanding that AI coding has eroded: two or more people working together with a single prompt/workflow.
Code, life, business, ruby, and AI.
I wonder if “shared prompting” could help counter the deep code understanding that AI coding has eroded: two or more people working together with a single prompt/workflow.
My typical workflow for larger tasks: research → interview me (grilling session) → generate a spec → tickets → implementation There is some iteration and review along the way. Once implementation is completed, there is a final deep review, testing, cleanup, etc. But recently, I have started to work in an early prototype or two. Prototypes used to be expensive. You would build something that just…
My model breakdown: Fable - the only reason I use Anthropic models for anything. I typically use it for planning. Sol/Luna - the workhorses. Luna is underrated, especially on cost. Various one-offs with OpenCode Go (which is likely the best $10 you can spend).
We should be prototyping more. Plans, specs, etc. are helpful, but there’s nothing like actually trying something out. The investment in a prototype used to be substantial. Now you can just kick one off and give it a try. Any lessons or shortcomings can be fed back into the spec. You can, of course, iterate on a “one shot,” but the end result is rarely as clean as if it had been done right the…
Issue #3 of AI AND Me is out Herdr over Solo, don’t pick a single model, background agents, skills, and my new AI sleeping giant - CloudFlare https://scottw.com/ai-and-me-issue-2026-3/
Lewp was upgraded to 0.2.0 - It now handles releasing of ports when the directory no longer exists (ie, you deleted the worktree or Stooges workspace before you released the ports)
Between life and the constant changes in AI engineering, this post has been nearly a month in the making. Things change, and we adapt so quickly that it is hard to sit down and think through what is actually going on. Here is what my current stack and setup look like. The Agent Runtime A year ago, I would not have given agent runtimes much thought. Once you start kicking off more tasks, including…
Automatically create @herdrdev workspaces when you add a workspace to http://Stooges.dev - https://gist.github.com/scottwater/9467447340c62875d955b50d16e9e19c
Herdr + Pi is a very flexible combination. You can run them just about anywhere, with just about any model. Both have strong extension stories, so even if you have a very specific workflow in mind, you can likely modify one or both to meet your goals. This is the way!
I swear, if I see one more video or post that says, “just ask your agent what three things to build next”… Never do this. Don’t build what your agent wants you to build. Only build what your customers need (not even what they want). Figuring out what they need isn’t simple, but building arbitrary features is a recipe for a crappy, bloated product.
If you’re ever curious how a new tool might fit into your workflow, give @mattpocockuk ’s Teach skill a whirl. It’s a fantastic way to go from zero to functional fast. Many tools come with a quick start, but this path keeps it focused on your goals and needs. You will likely learn something new about tools you already use every day.
I took a few minutes to finally connect Stooges → Lewp → Herdr. Now one stooges b feature-1 command sets up an isolated workspace in Stooges (and Herdr), gives it a local SSL URL http://app.feature-1.local.kickofflabs.com , and completely isolates PG and Redis databases. This works because Lewp can manage ports for more than just the loopback. A full blog post and sample app are on the to-do list.…
Not a bad week. Could have used a bit more zone 5, but the goal is consistency, not perfection.
Another reminder that you can do far more with AI than build features and apps nobody asked for. Burn your dependencies - https://scottw.com/burn-your-dependencies/
Impeccable by @pbakaus has been extremely helpful in spotting UI issues in my various projects (both agent-driven and self-inflicted). I highly recommend checking it out and letting it run a quick audit. https://impeccable.style/
There are people on X talking about canceling all of their SaaS subscriptions and building their own personal versions of those services. I find this silly. Building is easy. Maintaining is hard. Maintaining things you don’t really care about is a recipe for disaster. Having said that, from a developer perspective, I do think now is the time to look at your gems, npm packages, etc., and figure out…
PlanetScale - It would be great if this remembered my previous settings or had some sort of templates. I love this feature, but I would rather not have to click through and make the same selections over and over.
If you use Pi and SoloTerm, I made quite a few fixes and updates to the pi-soloterm-tools extension (basically the SoloTerm MCP for Pi). In addition, the Solo skill received a lot of polish. https://github.com/scottwater/pi-soloterm-tools
I have been using @mattpocockuk ’s skills quite a bit the last few days, and batch grill me has made a massive improvement. The skills do a good job of telling you what to do next, usually with a fresh context/session. One interesting area to explore would be automating this next step with Soloterm since you can easily get a fresh context and keep the entire process moving forward.
All I want from the next version of MacOS is one single security prompt per application. Allow the app to request all the access it needs and then give me one modal to accept/deny/etc. The security theater loop does not help anyone.
Debugging with the Sentry MCP/CLI along with the PlanetScale MCP is a massive superpower. On top of that, PlanetScale’s insights + recommendations find issues before they become a problem.
Slack needs a “move message to thread” feature or a warning, “are you sure you didn’t meant to reply to. thread” warnig.
Not having Fable available in Pi as part of my subscription is one of the more frustrating parts of my workflow. Everything I do (with Solo) becomes much easier to control if I can assume Pi is the harness.
KickoffLabs has been running on PlanetScale for about four weeks now, and the process has been super smooth. Props to @jessethanley for the indirect push. The tools and insights are so far ahead of Heroku that it now feels criminal to have waited this long. My only minor complaint would be that I hate having to think about egress. It has been minor, but it was nice to just ignore this with Heroku.…
One of the biggest mistakes you can make today with AI is going all in on a single harness, tool, provider, etc. Harnesses, tools, and providers are changing too quickly to reasonably commit to any one of them at this time. You have to keep reminding yourself that we are in the infancy stages of AI. It is unlikely that anything you use today will still be common practice a few years from now. With…
I set up the pi-grok-cli in Pi, which provides adds grok and the rest of the cursor models/etc to Pi. All that is missing now is some access to BugBot. https://pi.dev/packages/pi-grok-cli
Let’s go, SoloTerm style.
When coding with AI, take advantage of the opportunity. Look for ways to be more ambitious. Doing more, faster is overrated. Find ways to do it better. Copying and pasting a linear ticket into a prompt will eventually be 100% automated (heck, even tickets may become a relic of the past). There has never been a better time to take things a step further than what’s expected.
The only model decision I know to be 100% accurate is this: don’t rely solely on a single model. When you make important changes, updates, or other significant decisions, seek a second (or third) opinion, just like you would in real life. Some friction here and there is a good thing. Use it to learn for yourself so you can spot the weird things more easily in the future.
You can create a lot of new things with AI. It’s fun and gets most of the headlines. But you can also do a lot to fix the sins of the past: kill unnecessary dependencies, inline third-party dependencies, and remove unused dependencies. The calculus of tackling code debt has changed. Take full advantage of this opportunity.
Superpowers with Fable has been interesting. It appears to consistently be delegating implementation to Sonnet and Haiku, which is exactly what you would want to see. I may work on my SoloTerm version to do similar but across model providers.
I went all in with @sentry this week. My only regret is waiting this long. One data source for everything that is important to monitor. 💙
It is surprisingly difficult to get data out of HealthKit and have it match what Apple presents.
Switched my voice-to-text app to FluidVoice this week. Handles spoken edits and formatting changes smoothly, even mid-thought Near real-time’ish preview Great one-key activation toggle https://altic.dev/fluid
Updated TestFlight build for http://muv.ing/ went out this morning. It ships with a brand new overview of what you did during the day/week/etc. and focuses much less on individual activities. I ❤️ this change and feel like it really makes this app pop.
I built a new thing, lewp (loop). Lewp is a different take on the puma-dev experience and should pair nicely with the kontainer project I have also been working on. Think process isolation, limited blast radius, and no port weirdness. Plus it handles setups that require multiple ports (such like HMR/websockets). I need to kick the tires a bit more, but it should be available soonish.
A simple way to make using various models easier with SoloTerm - Just create new agents for each model.
GML 5.2 via @AmpCode has been great. It has taken some time to get used to the responsiveness (good problem).
I hate when console logging (puts in Ruby) is visible while writing specs.
Scratchpads in @aarondfrancis Solo are a hidden gem. Giving agents a simple, shared place to share state and provide feedback for the huma is fantastic. Now, if we could just add a bit of styling/theming for said humans…
I had an idea last night to clean up the Muv activities screen a bit more. My goal has been not simply to rehash a list of activities you can easily find in the Fitness app (and others). The second image is the updated summary view of every way you moved today.
PSA: hey dads. No middle-aged father of three dies from a sudden heart attack. It was never sudden. The damage has been accumulating for years. So get out there and move. Lift something heavy now and then. Get your heart rate up a couple times a week. Take the stairs. Park at the back of the lot. Walk… a lot. That’s the itch I’ve been trying to scratch. I’ve been building an app, Muv, that gives…
Claude Design is far from perfect, but it does the best job of any tool I have tried to visualize a wide variety of design choices. The scale is way off in this photo, but it makes the next decision much more obvious. B probably has the best feel and usability, but I may add an experimental feature to try out C (swipe from day to day, week to week).
Cleaned up the UI a bit. I am still not loving the next/previous navigation in the middle. I am thinking of dropping the “home” in the middle and moving it to the far left/right. I could also drop them entirely and support swiping to move from day to day (week to week).
CloudFlare’s email forwarding is magical.
I typically try to will a .com domain, but I am quite happy with the .ing
Muv is an app that I always wished existed, but pre-AI coding, I could never justify the time to make it happen. Zero business aspirations here; it is something I just always wished existed. https://muv.ing/ - TestFlight should officially be available in a day or two.
Almost had a regression in the move from Heroku Redis to DragonFly with a Lua script. Thankfully, younger me had this feature well tested, so older me did not fall on his face. Even better, older me can now see the Lua script as unnecessary complexity. Yay me.
TIL - DragonFly (Redis alternative) caps keys returned in a query to 8192. I should have been using dbsize to verify the number of keys in a migration anyway. Old habit.
As I hinted at in the last AI & Me , the biggest, newest model should not run every step of an agent workflow. The obvious reason is cost. Why pay two to ten times more than you have to? The less obvious reason is time. Fable (and whatever follows it) is slower. Time, even at the speed of AI development, is still valuable. An expensive and slow model still belongs in your workflow, but you need to…