sharing a skill i just published - it's called "/headless" it does one single thing - help you set up a mac, pc or linux box as an always-on headless machine if you have machines sitting around, try this out and put them to work!

WTF?? did X just rug-pull on the no-ads subscription feature? i pay the $200/month subscription and it’s suddenly ads like this everywhere a few weeks after the new head of product joining and now this??

before writing a lot of code with agents, first ask your agent to create a lint rule that caps how many lines a single code file can have. i typically go with 1000 this is the easiest win that has a lasting impact on the health of your codebase and do it as early as possible in your project, because unwinding it is costly later on

my open source projects collectively crossed 30k+ stars on github, going from almost zero just a few months ago! 🎉 reflecting on the journey, here are some learnings about growing and maintaining open source projects: 1. make videos sharing your OSS projects videos, not tweets. this is probably my most important tip i had a few viral tweets introducing my OSS projects that got lucky and had hundreds of thousands of views, but what brought in way more serious users and contributors was my youtube videos good videos take a lot more effort to make, but it gives people a much more direct understanding of how your project works and how it delivers value. my hero project firstmate grew largely due to my video content make some videos. and make some more 2. use open source as a super power i can't overstate how much better the tools got because of open source contributions from others. people bring great ideas, and test + fix the product in all kinds of edge cases that i can't possibly discover in a way, an OSS product can guide itself towards PMF, because the community would collectively steer it towards the right direction now that even non-developers can participate in contributing to OSS, this super power is becoming even stronger 3. set a minimum bar on contribution quality the super power can also be a double edged sword here. especially with AI which allowed a lot more people to contribute but also lowered the bar. if left unchecked, you get a lot of noise some projects outright disabled all PRs, while some others refuse to take anything done with AI. neither of which is the right approach IMO my solution has been to raise the minimum bar by requiring all contributions to be raised through no-mistakes. it enforces a level of rigor, and slightly increases the cost to contribute so low quality noise would often shy away i have a github action for enforcing this - which you can use for your projects too 4. automate as much maintenance as possible if you've contributed to my repos, you might have seen some of my automation at work. here's an example - i receive dozens, sometimes hundreds of incoming issues + PRs every day across my repos - i can't imagine how i could have kept all these projects moving if i hadn't leveraged agents to automate the triage work i run the automation in grok bot with my firstmate bot there - it can set up a different bot per each repo, and creates the routines to triage incoming issues and PRs using cursor cloud agents according to my rules, yet allows me to talk to just the firstmate about everything an important pre-requisite to doing that is a good documented vision for your project, which i write using - with a clearly written vision statement, your triage agents can automatically rule out changes that don't align, and focus on the ones that would further the project towards the direction you want

working with AI is fundamentally asking someone else to produce the work you want we spend a lot of time communicating context, goals, giving feedback, setting rewards, shaping workflows and SOPs. it’s a managerial experience i’ve always felt a bit sad about that, because it took away the experience and joy of being a pure craftsman that produces great work directly chopin can’t be chopin by hiring a team and telling his employees to compose for him. these masterpieces had to be created by a craftsman directly, and i think many of us humans deep down seek that mastery and joy when the world gets filled with more and more agents and robots, every human’s job becomes an agent manager are we sure that’s what we want?

just published a free OSS tool "runner-up" which is the setup i've been using to self-host all my github actions runners across my own machines i run thousands of CI jobs every day on private repos. they would have cost me about $5000-$9000/mo depending on what cloud provider i choose, which is completely bogus because that's a new mac studio every month after lots of calculations it was clear that simply owning my own hardware would be an order of magnitude more economical, which led to me creating this setup and i finally got some time to get it published if you have a similar need, hope this helps!

i’ve been trying to get rid of adversarial reviews for quite a while now, but unfortunately every time i stopped it, i would find myself simply delaying the discovery of real problems, and the later i find them the more they cost as a data point, i had measured how many real problems my adversarial reviewers actually caught, and it’s a lot.. that said, a good adversarial review setup is necessary for the ROI to be justifiable. i think a few things are very important: 1. original user intent must be explicitly collected (usually can be found in session transcripts where user inputs are) and passed to the reviewers 2. there needs to be a human in the loop for ambiguous decisions. any scope change / expansion needs to escalated to the human user for decision 3. reviewers need to treat complexity like a disease and actively point out simplification opportunities 4. reviewers need to challenge the lack of empirical evidence and hold the implementer accountable to prove their change

i’ve been trying to get rid of adversarial reviews for quite a while now, but unfortunately every time i stopped it, i would find myself simply delaying the discovery of real problems, and the later i find them the more they cost as a data point, i had measured how many real problems my adversarial reviewers actually caught, and it’s a lot.. that said, a good adversarial review setup is necessary for the ROI to be justifiable. i think a few things are very important: 1. original user intent must be explicitly collected (usually can be found in session transcripts where user inputs are) and passed to the reviewers 2. there needs to be a human in the loop for ambiguous decisions. any scope change / expansion needs to escalated to the human user for decision 3. reviewers need to treat complexity like a disease and actively point out simplification opportunities 4. reviewers need to challenge the lack of empirical evidence and hold the implementer accountable to prove their change

first IDE fell apart, then code review became broken, CI became a bottleneck, ands merge queues stopped working. the point of failure kept moving “right” are you ready? because the next stop is production, incidents and oncall

def not a small thing - fast and cost efficient ci/cd is a big deal especially as our velocity ramps up, and it's not easy for full stack cross platform products if you are on a team that has a slow pipeline, getting it optimized is a very impactful thing you can do

For 9 weeks, I mostly talked to one AI chat. ⠀ In that time, it finished 560+ tasks and merged 365 pull requests across my projects. ⠀ Yesterday, at Builder Sundays hosted at @Shopify's Ottawa office, I showed how. The tool is @myfirstmate, a free, open-source project by @kunchenguid, a former L8 engineer at Atlassian. ⠀ Kun built it because juggling parallel agent sessions was killing his focus. Every switch meant remembering what that session was about and what to do next. He didn't want a smarter workflow. He wanted the calm back, and the confidence that nothing falls through the cracks when you look away. ⠀ How it works: you are the captain, one agent acts as your firstmate, and a crew of agents does the work. Nothing goes live without your yes. ⠀ → Orchestration. I only talk to firstmate. It breaks the work into separate tasks, starts a dedicated agent session for each one, and runs them in parallel. It keeps an eye on every session, handles the small stuff itself, and only comes to me when a decision actually needs me. When a task is done, it brings the result back for my review and approval. ⠀ → Memory. It keeps my projects, rules and in-flight work. /bearings is my morning brief, /stow saves what it learned. I close the laptop and pick up tomorrow where I left off. ⠀ → It improves itself. If something isn't working, I ask firstmate, and it fixes itself. ⠀ The result: Partexia went live on iOS and Android, and CrewTower shipped 23 releases. Thank you Kun Chen for building this and sharing it with everyone.

i just read through this article of deepseek’s super-agent it’s a good write-up describing how firstmate works 😄 only half joking - it’s indeed the same architecture and folks using firstmate have been enjoying SOTA for months

over the past couple of weeks, i heard multiple people telling me they could no longer use up their LLM subscriptions the reasons were a mix of: 1. the frontier models became more efficient 2. they felt exhausted managing agents 3. they ran out of problems that they can throw to AI 1 will continue to happen. 2 is a tooling and workflow problem (just use firstmate haha) that will be solved as well 3 is an interesting one - it points at a novelty effect that may be wearing off, and a capability gap that AI just hasn’t reached the level that most people can trust to run their daily lives yet meanwhile, my quotas have never been tighter because i have an endless number of things i want to build that points at the reality of user segmentation - R&D is such a niche use case that needs an astounding amount of tokens. there aren’t many other use cases that are so token intensive realistically, how many % of the world’s population will want to run an army of agents and build software all day? i don’t expect that to ever become a big number so my speculation is that the current $200 subscriptions are likely already close to the ceiling of most people’s consumption needs. better models will unlock more use cases, but efficiency gains may cancel that out like water, gas, power - most households will simply get one or two monthly AI subscription and be done now the important part - this means if you are building a B2C business, you should almost never sell tokens as part of your offering, because there will be no demand. people already have their subscriptions and that’s enough instead, allow people to bring their own LLM subscriptions to your product, and add value on top. i suspect this will hold for most B2B businesses as well, but for B2C this is almost certain at this point

ok it's official... AI-written unit tests and integration tests are empirically proven to be unhelpful. you should tell your agents to stop writing tests by themselves on the deepswe eval set, banning sonnet 5.5 high from writing any tests actually resulted in slightly higher success rate (non stat-sig), with less time and token spent (stat-sig). this is pretty hard empirical evidence across the tests that were written by the "tests allowed" baseline arm, 65% of them were unit tests, 35% were integration tests, neither bucket resulted in any improvement compared to not writing any tests at all i also picked a random sample of 44 tasks subset where i completely disabled executing even existing tests - it also did not affect success rate at all i spot checked many tests written in the baseline arm, and my intuition is that most tests are simply a repetition of the implementation agent-written tests do not add any value because both the implementation and the tests were simply the agent's interpretation of our intent. the tests aren't any more accurate than the implementation itself i suspect we can still extract some value from unit tests and integration tests if we describe them ourselves when we believe we can articulate our intent better through test cases than through requirements. but i have not proven this yet also worth noting, during deepswe eval the agent wrote almost no e2e tests (only 17, compared to 3000+ unit/integration tests written). so this analysis does not prove nor disprove the value of e2e tests - i will do another eval specifically for that

AI today can easily create pictures that mimic van gogh’s style. but.. can it create a new van gogh? learning about van gogh’s life made me realize the way we are building super intelligence today is fundamentally limiting van gogh’s unique style came from his unique journey, which is a result of having a limited experience of this world it’s the limit of our lifetime that forced every human being to walk a different path, and that’s what created unlimited differences feeding a super intelligence all the data in the world is exactly what ended up limiting its ability to have a truly unique perspective and in a way, feeding the same super intelligence to every human being will propagate the same risk to us all, and over time can make us lose the creativity we used to have perhaps, the most human thing we can do is to go out and experience the world through our own lens, get inspired, and create something truly unique with the gift of a limited life