Not sure that Claude likes using Skills.
I have a pull-request skill that nearly never triggers even when using /pull-request as the command.
Only explicit requests seem to work.
Everyone is talking about software factories.
I'm all about software gardens.
Bought my first .ai
The fibre backbone is solid from what I understand.
We don't have issues with earthquakes.
We should be building chip fabs and data centres as fast as our skilled trades people can assemble them.
@build_canada what do we need to do to make this a reality?
Why on earth are we building data centres in Texas and Tennesse to power AI when they are very warm places?
Average temperature in Canada is much cooler. There is space beyond measure. There is natural gas, coal, and even uranium beyond measure. Fuel for power generation is not an issue.
There is boatloads of UX work to be done in the AI-era. Some gnarly problems to solve. If you love UX, dig in, use the tools, and you'll see a myriad of opportunities to ply your trade.
Man, I love @elixirlang. The deeper I get into the ecosystem the happier I become. It's so simple and so powerful. People are really sleeping on it, I think.
Making software is much more like gardening than running a factory.
I think making software is more like gardening than running a factory.
Factories make goods based on a spec. The goal is zero variation from the spec.
Tending a garden means responding to what the garden is telling you about what it needs.
Payments companies will need to adapt to this new reality or they will lose to crypto simply because it can be used by an agent.
With the rise of agents, crypto based transactions are going to become the default.
Have an agent make a payment is much harder with a credit card.
There's no CLI equivalent of the iframe credit card form or checkout page.
Rest Day —
“The LORD your God is in your midst, a mighty one who will save; he will rejoice over you with gladness; he will quiet you by his love; he will exult over you with loud singing.”
Zephaniah 3:17 ESV
You remember when World of War Craft was released and how people got fired and were found having not left their house in days....
ya...
This is where we're at with AI right now.
A tiling window manager with multiple workspaces might become a necessity in this new highly parallel world.
@OmarchyLinux I'm looking at you. 🤩
I think this means ultimately you need to combine strong determinism with the agent's agency.
Not everything should be done by an agent.
The goal is efficiency and quality so that means choosing the right tool for the job.
But even then this just pushes the human review up a layer, like we would push it up the leadership structure.
If agents are building for humans, at some point a human has to decide what's good and be able to thread feedback through the system.
We will have to develop ways to define this meta work in a computable manner.
We can demand more of the LLM. Let the agent check its own work. Yet at the end of the day, what's good and not good is up to the people who are building the product.
Cognitive load, I predict, will become one of the dominant topics going forward.
How do we manage the cognitive load of having so many parallel streams of work coming in. Author vs Air Traffic Controller
This matters because at the end of the day the tool needs to serve me and my efforts to deliver a quality product.
If everything is different every time in subtle ways, that's cognitive load.
I want high agency and intelligence in the solving of problems.
And I want legalistic discipline in following the project's workflow.
Chaos is good in the micro.
High predictability is need in the macro.
It mostly gets it on the second try, but this lack of reliability is annoying. If we're building software factories, the lack of determinism in the assembly line is moving in the opposite direction.
I take this to mean that I need tools to enforce the workflow.
I'm all for the agent doing its thing in solving the problem. Be a high agency agent.
However, not following the project's workflow requirements reliably causes more work for me than I want. And it's dumb work too. "Use pull request skill to make the PR." "Update PR using pull request skill."
As I push to do more and more work with as little intervention as possible, the main piece of friction I'm noticing is that the models don't follow workflow instructions reliably.
I have commit and pull request workflows outlined in Skills and reference them. They are only sometimes used.
Don't surrender agency to the model. It only knows the past.
If you let the model lead your creativity, you will get something that's derivative of the sum of past human knowledge.
If you use the model to augment your creativity, you will get something that leverages the sum of all human knowledge to elevate your creativity.
The more I do this agentic programming thing, the more I'm convinced that agent session transcripts are a valuable artifact to the completion of a task on par with the code generated, the design work, and any other supporting product documentation.
As multi-agent workflows explode in popularity, people are going to struggle to comprehend what's up. The best tools will put the needs of human cognition at the centre of the design and support the management of attention.
Overwhelm.
UX to combat overwhelm is something you're about to hear a bunch about.
With an agentic workflow where PR volume is way up, this approach drives down push to merge latency.
