This is post 2 of AI-Powered Product Development: From Idea to Company with AI. Post 1 is here.

Read More

The exhaustion pushes you to go along with whatever direction an agent is taking, and you keep rolling the dice, hoping the end result will be in your favor. This has gotten me into trouble at times.

My history over the last year has been one of pushing agents and AI to do more as their capabilities expand. I reached a point several months ago where I started to build on different projects, and each of those projects usually has more than one agent session running. If you work somewhere that requires close review of every line of generated code, you are probably not running sessions in parallel like this, and some of what follows will not apply.

Read More

I've sat through several vendor evaluations over a long career in technology. This one was different in two ways. It was a workshop, our first contact with the vendor, rather than a formal evaluation. And it was the first time the product was an AI system.

I had very few technical details going in. What I understood was that we were looking at an intelligence management system, a sort of fancy RAG setup.

One of the first things the vendor team demoed was the system finding the gaps in our own knowledge base. Honestly, what they're selling has merit. It isn't just RAG, and it looks like a nice piece of technology. That's what started the trouble. A workshop has no scoring sheet, and nobody in the room was holding a checklist. Because the thing had real potential, my colleagues began exploring what else it could do, and within a few minutes we were talking about an AI hub that would answer all our prayers, a missing piece of our AI stack finally falling into place.

Read More

Before the post, a word about how it was written

This post was written mostly by an AI. I want to say that up front, because I keep seeing outlets ban AI-generated content outright, and I think that's a mistake.

I'm an engineer, not a writer. Writing takes me real effort. Most of the time I do it anyway because I enjoy it and because it forces me to think a topic through properly, which is its own payoff. It brings me clarity. But that's a few hours I don't always have.

So the choice in front of me this week wasn't "good post or AI post." It was "AI post or no post." I found something genuinely useful, almost by accident, and I've seen it mentioned in exactly one newsletter. Nowhere else. If the price of sharing that is admitting a model helped me put the sentences in order, fine.

The blanket ban treats "AI-generated" as a proxy for "worthless." Sometimes it is. There's a lot of garbage being pumped out for clicks and I'm not defending any of it. But the proxy is bad. What matters is whether there's something real underneath: a discovery, a bug you actually hit, a thing you actually built. Judge that. I'm not doing this for engagement numbers. I'm doing it because I think three or four people reading this will go install this thing and be glad they did.

Anyway. On to the important part.

Read More

I built this for me

Eight newsletters in the inbox, checked every evening. That was my N. I was wasting too much time reading the same story multiple times and having to blow past low-value material. There was no moment of inspiration, no interesting backstory (although the AI editing this piece tried to coax me in that direction). I had a repetitive reading chore, I can write code, and the chore happens to be a use case that fits what a language model is good at if you build enough scaffolding around it to make the output trustworthy.

Read More

Seven years ago, I was sitting at a bar after work, having a quick beer to relax before heading home for my second shift with my newborn son. Or maybe at that point he wasn't so newborn, but he was still a baby. An older lady sat next to me with a dog. The dog was very docile, probably an older dog, just lying on the floor, waiting for its master to finish her drink. When the lady got up to walk the dog out, I saw how the leash had become tangled around the dog, and she spent a minute untangling the mess. Watching that, I came up with the idea for a better dog harness. For the next few months, maybe a year, I worked on the idea on and off, but then it died because the next hurdle (actually creating a prototype and filing a patent) was just too much to get through after dealing with all the demands of day‑to‑day life.

But I didn't let it go. Not entirely.

Read More

How I went about securing Hermes Agent, an always-on CLI-based AI assistant from Nous Research, running on a spare mini PC at home. This is not a "here's the perfect setup" post. It's the honest record of three approaches that failed and the one that triumphed.

Read More

The category

Both projects are self-hosted personal agents: a persistent process on hardware you own, wired into your messaging apps, with real tools (shell, files, web, a browser), memory that survives restarts, and a skill library that grows. You message it from your phone; it acts on your machines.

That last part is why security dominates the comparison. A chatbot that answers wrong wastes your time. An agent with shell access that gets manipulated does real damage.

Read More

Two tracks, and why we started with people

Bringing AI into an organization runs on two tracks. One is top-down: leadership and departments looking at processes and finding where AI can improve them. The other is bottom-up: helping individual people do their own day-to-day work more efficiently. You start with drafting, summarizing, translating, working through a document or a spreadsheet, then graduate to creating personal agents that optimize a very specific workflow. The stretch goal is to make everyone a builder, or at least give them the power to be.

The process track matters, but in many organizations, ours included, a lot of it depends on data that isn't yet clean or well connected, and that work takes time. The projects that need that data wait on it, while others move ahead. The individual track doesn't have that dependency, so it can deliver value sooner. What it needs is smaller: a usable tool everyone has, and some guidance on how to use it.

So we started with people. And before you can help anyone's everyday work with AI, they need access, which made the first step concrete: give everyone a single tool they can all use, good enough to be useful, running on infrastructure we control, with data kept in the EU. The rest of this post is how we built that.

Read More

The record you don't otherwise have

A while back I wrote a breathless post about building this pipeline, turning permissions off, and one-shotting features. I stand by the thrill. This is the sober follow-up, and it starts with a boring fact: when an agent runs with permissions off, you are not watching. That's the point of it and the whole problem.

Read More

The problem

I built MemodoAI, a fork of LibreChat, to give the whole company an AI chatbot and agents without per-seat subscriptions. Then I had Claude produce an 18-minute onboarding tutorial. Both helped. Neither fixed the underlying problem, which I'd seen well before MemodoAI: most people use AI tools for a narrow slice of what they can do. An email rewrite, a summary, and that's it.

Access isn't the bottleneck. Instructions aren't either. The missing piece is the link between what a tool can do and the specific work a given person does every day. A tutorial covers the features. It doesn't tell a controller which of those features applies to their month-end close.

Read More

I thought that developing and rolling out an internal AI assistant would be the impressive part. Turns out that the tutorial and user guide created mostly by Claude is what blew me away.

Claude clicked through every screen of our internal application, typed the demo prompts, narrated every section in a voice that wasn't mine, built the branded title cards, and transcribed the whole thing in two languages. For the length of the shoot I mostly sat there and watched my own app give itself a tour.

Read More

Realization

I've been using Claude and Claude Code for a while now, and I've sat through several model updates without much friction. Opus 4.7 was different. Over the last few weeks I'd noticed I was having to work harder to digest what Claude Code was telling me. Not the code, but the analysis, the progress reports, the wrap-ups at the end of a task. They'd gotten terse. The harder the project, the harder it got to follow. So naturally, I asked Claude for help.

Read More

A Giant Pile of Shit

I thought I had mastered spec-driven development with Claude Code, and more fundamentally become an agentic coding superhero. That was until my ambitions overtook capacity. Agentic coding has gotten exceedingly powerful, but it still requires a developer to maintain close supervision, especially in more complicated scenarios.

Around six months ago I built this SDD development workflow plugin for applying a research-plan-implement cycle to software projects. Learning from my experience using it, I continued to tweak it and added additional review phases and steps that auto-corrected imperfections. It worked so well that I figured I could more fully automate the process, which led me to create the /sdd-flow skill. This allowed me to simply one-shot features (I am using "one-shot" loosely here because there are clarification and review steps involved). That worked well until my projects got complicated and one-shotting meant handling too many requirements and expectations at once. In short, I had drunk the Kool-Aid and believed AI could deliver.

Read More

So working with an AI agent has been pretty revolutionary for me. It has allowed me to automate and run jobs and workflows that would have normally occupied most of my time, now in a much simpler manner, independent of my time invested. This isn't news; I think everybody sees this experience shared by many online. What I'm seeing more and more now is the ease or the accessible ability to coordinate many agents that then obviously multiplies your effectiveness. This coordination of agents is happening in the spaces of software development, of personal agents, and even in a more abstract way within organizations.

I don't mean to add to the pressure that is very palpable for many of us, where there is a sense that you are being left behind. I do want to encourage everybody, with a strong sense of urgency, to start engaging with these tools, because it takes some time to truly understand the shift in possibilities and to understand the realm of productive environments that are now available with this technology.

Read More

"Sales cures all ills" is a phrase one of my first mentors repeated often, and it stuck with me. Before I got serious about software development and before meddling with AI engineering, I was a salesperson. I still miss the thrill of closing a deal and outmaneuvering your opponent. It was fun. Now that I'm in a position to use AI at my company, I came up with an idea: use voice-enabled AI to help train and coach our salespeople.

Read More

We are all energy. Everything is energy. All these energies interact with each other. This is not an all-encompassing description of how this universe behaves, but it's an accurate one.

I got to the topic of energy because I was trying to fight the desperation I feel as I consume news that reveals all the depraved acts described in the Epstein files. As I watch innocent children murdered, dismembered, disfigured, and traumatized as a result of the violent acts of greedy and misguided men. Starting wars as easily as we choose our outfits for the day. Maybe I was just not conscious of this earlier in my life. Maybe now that I have kids, I am more sensitive to all of it. But at no point in my life have I felt more certain that evil is winning.

Read More

Last week I shipped Redakt, an open-source GDPR compliance tool. Six features for a web UI and REST API to help us better handle PII (Personally Identifiable Information).

I had been using the spec driven development (SDD) process for months through a plugin I built. When I saw that Claude had expanded the context window to a million tokens, I thought to myself, "Why not try to push the limits of how I was developing software?" What I mean by that is that the SDD process was created to help with context management. It follows a traditional and rigorous software development cycle, but it was slow relative to how fast things move within the AI space. I decided to try just one-shotting a feature using a skill to orchestrate the phases of a development cycle, and it worked surprisingly well.

The skill was not the only thing that optimized the process. Anthropic's release of --permission-mode auto last week was the push that I needed to remove the diapers. I was excited to try it out. This mode is limited for now to Teams accounts. It was flaky initially, but it gave me a taste of the forbidden fruit. I didn't want to go back to confirming every request. At home, I knew there was another option though, a dangerous one, that I had avoided until now: --dangerously-skip-permissions. After close to a year of using Claude, and never having had any "oh shit" moments, I decided to remove all guardrails and bypass permissions. It felt like my face was being pushed back by gravity as you see in fighter pilot videos when they fly at multiple Machs... and the feeling is addictive.

Read More

Earlier this month I wrote about shadow AI and the compliance gap discussing how employees using unapproved AI tools with personal data are creating quiet GDPR liability across Europe, and how the gap between what the law requires and what companies actually do is growing every day.

The response made one thing clear: people know they have a problem. Current measures feel like we are all playing whack-a-mole.

The advice that post ended with was honest but incomplete. Department-specific guidelines, approved tool lists, clearer communication — all necessary, all insufficient. Because even with perfect policies, you still have the same fundamental problem: an employee sitting in front of ChatGPT with a paragraph of text containing a customer's name, email, and order history, and no practical way to strip it out before hitting enter.

So I built something.

Read More

My Bookmarks Saved Me Money

A few weeks ago, I was designing a multimedia publishing pipeline for this site. I needed voice synthesis and video generation tools, and I had a rough budget in mind, something like $56/month across a few SaaS subscriptions.

Before committing, I searched my Sift knowledge base. It surfaced two ComfyUI integrations I'd bookmarked a week earlier: Qwen3-TTS for local voice cloning and LTX-2.3 for open-source video generation with portrait support. I'd saved them as simple URL bookmarks with short descriptions — the lowest-effort form of ingestion possible.

Those two results shifted the entire pipeline from a paid SaaS stack to a $0 local approach.

Read More

A few weeks ago, as I started to write more about AI, I figured that what I was posting to the web should be easily accessible by the technology that I was writing about. I had Claude Code review my website structure and have it give me suggestions on how I could make it more AI agent-friendly.

I also added automated translations to three other languages in my publishing process. Beyond making the content more accessible to more people, it has the side benefit of surfacing in AI agent searches in those languages.

This morning I came across Julia Solorzano's post on answer engine optimization, which gave me some additional ideas. Her post approaches it from a slightly different angle, and that made me realize this information might be useful to others.

Read More

When Search Can't See What You Mean

In the first post, I described the moment when good organization still isn't enough. You can tag perfectly and still not find what you need, because you're searching for a concept and your note used different words. That's the ceiling of keyword search.

I have around 1,900 markdown files across three Obsidian vaults. If you're not an Obsidian user, you can just think of this as 1,900 text files in three different folders. They're tagged, linked, and structured with frontmatter, or more simply put, page properties. The organization is solid. But when I search for "authentication," I don't find my note about "session token management." When I search for "AI assistant," I miss the note about "Claude Desktop." The information is there. The search just can't see it.

Read More

Series: AI-Powered Knowledge Management — Post 1 of 9

The Folder Trap

I was a heavy OneNote user for years. It was perfect for me, at that time. I had nested sections inside sections inside notebooks, color-coded and carefully maintained. When I needed something, I navigated a hierarchy I'd built from memory. It worked as long as I remembered where I'd put things, and as long as the information fit neatly into one category.

It often failed to fit into just one category. This was frustrating, but I lived with it.

Read More

Why I'm Feeding an AI Every Piece of My Work Life — and What It Actually Needs to Function

A few weeks ago, I wrote a post titled I Am No Longer Needed, about how I had automated so much of my project management role as an AI lead that I could clearly see the shape of what was coming next. That post was uncomfortably honest about the emotional weight of seeing so much change so unexpectedly quickly.

But I realized afterward that it didn't fully answer the question underneath all of it: why? Why am I so obsessed with capturing and digitizing information? Why does it matter that my meeting notes, my task lists, my requirements gathering, and my ideas all exist in a queryable format? Why did I spend months building a personal knowledge base?

This post is that answer.

Read More

A colleague and I were talking about Markdown files and how they've become the default format for working in AI environments. Naturally the conversation drifted to Obsidian because I am a big fan. He mentioned something that stuck with me: he couldn't get comfortable with the vault concept, Obsidian's way of containing everything in a single managed folder, and it was holding him back from the tool entirely. He'd dabbled with this great note-taking system, but his actual work didn't live in it. His code was somewhere else. His documents were somewhere else. His Obsidian notes were somewhere else. Three separate places, no connection between them.

This is a problem almost every Obsidian user runs into, and most people just accept it. I had become so used to working in the Obsidian environment that I had become blind to this disconnect. Now that it was visible again, I didn't want to accept it.

Read More

I was a Microsoft OneNote user for a long time. More than ten years. I could not even begin to count how many notes I had built up in that time — meeting notes, project research, code snippets, random ideas. A decade of thinking, captured in a proprietary format on Microsoft's servers.

As a developer, I knew there was a principle about managing data well: keep content and configuration in the same file. Keep it portable. Don't let your tools hold your data hostage. I knew this principle. I applied it in code. And then I ignored it completely for my own notes, for over a decade.

I eventually made the switch — to Markdown files, and specifically to Obsidian. What surprised me was the timing. I moved before the current AI wave hit. I wasn't chasing a trend. I was just finally acting on something I already believed.

What I didn't anticipate was how right the decision would turn out to be, and for reasons that had nothing to do with personal productivity and everything to do with how AI agents actually work.

Read More

Every time an employee pastes a customer name into ChatGPT, runs a vendor contract through DeepL, or asks Copilot to summarize their inbox using a free or unapproved account, they are operating in the dark — using tools the company hasn't sanctioned, with data the company hasn't authorized, under poorly understood legal frameworks. This is shadow AI: a quiet, daily habit that has become one of the most underestimated compliance risks in Europe.

I came to understand this not as a lawyer, but as an AI engineer. About two months into rolling out AI tools and creating guidelines and policies at work, I realized we had a problem: the guidance we'd worked hard to communicate to employees was missing something fundamental. Nobody had explained the difference between "confidential data" and "personal data" and the importance that this holds when using any of the easily accessible AI tools, and how this still applies even when working in a B2B market. Prior to the ChatGPT era, most of us found it sufficient to use standardized office applications to get our daily work done. Obviously, now our day-to-day work has changed significantly.

The policies had to be redrafted. Not updated — redrafted. We needed specific language covering GDPR requirements and how they overlapped with the EU AI Act, examples employees could actually apply, and clear distinctions between the tools they could use freely and the ones that required approved enterprise accounts. What made the gap visible wasn't an audit or an incident. It was the approval requests. As employees started coming to us to review, vet, and approve the AI tools they were already using — or wanted to use — a pattern emerged: most of them had no framework for thinking about what data they could share and with what. Shadow AI is a reality at most companies, and I can tell you that the liability it creates is not being addressed with nearly enough urgency — especially as free tools continue proliferating and competing for everyone's attention.

Read More

A few weeks ago, I published a post about automating my own role. I timed it deliberately — sharing it with colleagues before a casual lunch with management, hoping to spark some real conversation about what's happening with AI and what it means for our work.

The response was almost nothing.

A couple of colleagues joked nervously that it made them uncomfortable. Management said nothing. It's been two weeks, and still no substantial feedback. Everyone is busy, life is full, there's no shortage of terrible things demanding our attention. Maybe they haven't read it.

But the silence got me thinking. And what I landed on is something I've started calling the workplace blind spot.

Read More

Throughout a large portion of my life, I have been at the center of, or at least close to, issues that have shaped American society and are now rippling outward across the Western world.

This isn't a credential. It's a pattern I can't unsee.

The Bubbles

During my college years in the late 1990s and early 2000s, I was studying computer science and working side jobs as a web designer. Technologically, that meant having design skills and being able to string together HTML — simple stuff. But because I was inside that world, I could see what was coming. I remember the moment clearly: sitting around after smoking a joint, thinking about how much of a scam the dot-com bubble was. The details have faded, but the clarity hasn't. It was all so exaggerated, so fake. And it was all going to come crumbling down. The money crashed, but the technology lived on.

After college, I worked in what was essentially a boiler room — think The Wolf of Wall Street, but on a much smaller scale. I didn't make any real money there, but I spent several years getting a visceral introduction to the brutality of the financial markets and the type of people attracted to that environment. I also got a glimpse of the type of person you eventually become if you stay.

Read More

How I Automated My Own Role as AI Lead — and What That Actually Feels Like

This past week, I realized something uncomfortable: I had automated myself out of a significant portion of my job as a project manager. Not theoretically. Not as an exercise. I had actually done it — piece by piece, meeting by meeting, analysis by analysis — until I looked at what was left of my daily work and thought, this doesn't need me anymore.

I'm an AI lead at a mid-sized company in Germany. My job is to evaluate, implement, and manage AI solutions across the business. The irony is not lost on me that the tools I was hired to champion are the same tools that are now doing chunks of my job better than I was.

Let me be honest: AI has not done everything perfectly. There have been failures and outputs that needed heavy editing. But those failures are small compared to the sheer volume and quality of what it has delivered. The balance sheet tips overwhelmingly toward the machine.

Here's how it happened.

Read More

Source: A Guide to Which AI to Use in the Agentic Era by Ethan Mollick (Feb 18, 2026)

Practical Guide: Which AI Should You Use in February 2026?

The original article by Prof. Ethan Mollick goes much deeper into each of these topics with examples and comparisons — it's well worth reading in full — but here's a practical summary of the key takeaways.

The Big Picture

AI is no longer just chatbots. You used to type a question and get an answer. Now AI can actually do tasks for you — research, create spreadsheets, build presentations, organize files. To pick the right tool, you need to understand three simple concepts:

3 Things That Matter

Read More

Source: The "Innovator's Dilemma" of our society by DocIsInDaHouse (Feb 11, 2026)

Core Thesis

The article applies Clayton Christensen's "Innovator's Dilemma" framework — where successful companies fail by optimizing the present while missing disruptive shifts — to society as a whole in the age of AI.

The key problem: A-posteriori thinking. Drawing on Kant's distinction between a posteriori (knowledge from experience) and a priori (knowledge through reason), the author argues that our society, politics, and ethics boards operate almost exclusively reactively. We regulate social media only after it destabilizes democracies. We address data privacy only after massive abuse. This reactive approach worked when innovation was linear, but is fatal with exponential, disruptive technologies like AI.

Read More

Source: Software Factory by Simon Willison (Feb 7, 2026)

Simon Willison covers how StrongDM's AI team has implemented what Dan Shapiro calls the "Dark Factory" level of AI adoption — where no human writes or even reviews the code that coding agents produce. Their full writeup: Software Factories and the Agentic Moment.

Core Principles

StrongDM's AI team (founded July 2025, just 3 people) operates under radical constraints:

• Code must not be written by humans
• Code must not be reviewed by humans
• If you haven't spent at least $1,000 on tokens/day per engineer, your software factory has room for improvement

The catalyst: with Claude 3.5 Sonnet revision 2 (October 2024), long-horizon agentic coding workflows began to compound correctness rather than compound errors. The November 2025 inflection point (Claude Opus 4.5, GPT 5.2) further improved reliability.

Read More

Source: The Algorithmic Bridge by Alberto Romero (Feb 7, 2026)

Core Thesis

AI agents (Codex with GPT-5.3, Claude Code with Opus 4.6) collapse execution into wishing. The bottleneck shifts from how to do things to what to do. Most people have spent their entire lives optimizing for the "how" and are now holding a genie lamp while still spending 99% of their effort on execution.

Read More