<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:itunes="http://www.itunes.com/dtds/podcast-1.0.dtd" xmlns:googleplay="http://www.google.com/schemas/play-podcasts/1.0"><channel><title><![CDATA[cgardens]]></title><description><![CDATA[tech, intellectual honesty, etc]]></description><link>https://blog.cgardens.dev</link><image><url>https://blog.cgardens.dev/img/substack.png</url><title>cgardens</title><link>https://blog.cgardens.dev</link></image><generator>Substack</generator><lastBuildDate>Sun, 22 Nov 2026 13:06:29 GMT</lastBuildDate><atom:link href="https://blog.cgardens.dev/feed" rel="self" type="application/rss+xml"/><copyright><![CDATA[charles]]></copyright><language><![CDATA[en]]></language><webMaster><![CDATA[cgardens@substack.com]]></webMaster><itunes:owner><itunes:email><![CDATA[cgardens@substack.com]]></itunes:email><itunes:name><![CDATA[charles]]></itunes:name></itunes:owner><itunes:author><![CDATA[charles]]></itunes:author><googleplay:owner><![CDATA[cgardens@substack.com]]></googleplay:owner><googleplay:email><![CDATA[cgardens@substack.com]]></googleplay:email><googleplay:author><![CDATA[charles]]></googleplay:author><itunes:block><![CDATA[Yes]]></itunes:block><item><title><![CDATA[Holding AI agents accountable]]></title><description><![CDATA[Because it&#8217;s not your agent who is going to get fired]]></description><link>https://blog.cgardens.dev/p/holding-ai-agents-accountable</link><guid isPermaLink="false">https://blog.cgardens.dev/p/holding-ai-agents-accountable</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Wed, 04 Nov 2026 01:02:57 GMT</pubDate><content:encoded><![CDATA[<p><em><a href="https://lnkd.in/p/dtKXmgGQ">Originally posted on LinkedIn.</a></em></p><p></p><p>There's a lot of discourse about whether AI agents "care" about us, and I think it's mostly the wrong question. Not because the question is uninteresting (it's very interesting!) but because it doesn't really bear on the practical problem anyone deploying these things actually faces, which is: how do you make an agent accountable?</p><p></p><p>It's worth stepping back and noticing that this is a hard problem for humans too. If you run an organization with 40,000 employees, some of them are going to do the wrong thing. A few will be genuinely bad actors, but that's honestly not the main issue. The main issue is that perfectly well-meaning people, at scale, will skip the double check or tell the customer what the customer wants to hear. If that sounds a lot like your chatbot, well, yes.</p><p></p><p>Big institutions have developed a lot of tools for dealing with this. Some are blunt: you can lose your certification, you can get fired. But a lot of the work is done by culture, and the culture is very domain-specific. Pilots use checklists. Surgical teams do a timeout before the first incision. Banks require a second approver on wire transfers. None of this is glamorous, but it works reasonably well, and it works because it changes incentives.</p><p></p><p>The problem is that none of these mechanisms exist inside an LLM. You can tell it to make no mistakes as many times as you want, and it will still sometimes give you a wrong answer with total confidence. People hedge when they're unsure. Models mostly don't, and the APIs don't give you a reliable confidence signal to work with.</p><p></p><p>So the accountability has to live in the harness: requiring receipts for every output, having a separate verifier check the work, running the same task several times and treating disagreement as a red flag, and setting things up so that asking for help is the rewarded behavior.</p><p></p><p>A well-built harness can actually give you a stronger guarantee than human process. Human reviewers get tired; a verifier in the harness checks every output, every time.</p><p></p><p>But to be clear, this is a lot of work, and you have to do it workflow by workflow. It would be much better if the models just knew when they were unsure and said so. Until they do, accountability is something you have to build, not something you get for free.</p>]]></content:encoded></item><item><title><![CDATA[What to Get Right When Building a Software Factory]]></title><description><![CDATA[Because the agent did exactly what we told it to, and that was the problem]]></description><link>https://blog.cgardens.dev/p/what-to-get-right-when-building-a</link><guid isPermaLink="false">https://blog.cgardens.dev/p/what-to-get-right-when-building-a</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Sat, 31 Oct 2026 12:31:02 GMT</pubDate><content:encoded><![CDATA[<p><em><a href="https://lnkd.in/p/gd5WXKCE">Originally posted on LinkedIn.</a></em></p><p></p><p>If you're building a software factory, here are two steps that are easy to miss. We've been running one at Levelpath since April, and it has increased our development speed by more than 10x without sacrificing quality. Most people focus on whether the agent can write code. That matters, but the two things that held us back were harder to see.</p><p></p><p>1. Verification. Until you know a change will do what it was asked without breaking something else, the AI skeptics are right: shipping broken code at machine speed is worse than shipping slowly. That's why agentic QA and testing is the most expensive stage of our pipeline. What it buys us is a quality guarantee, the same way a physical factory stands behind what comes off its line. Once you can make that guarantee, you can put your foot on the gas. What that guarantee should cover looks different from the pre-AI bar, but that deserves its own post.</p><p></p><p>2. Input quality. We started with a ticket quality agent that checked for enough context and clear requirements. That was table stakes. But when we traced back the times we were unhappy with the factory's output, the agent had usually done exactly what it was told. We had told it to do the wrong thing. So the input stage grew. It now runs architectural and product checks, and the agent writes a spec before building, pulling in a human when it needs input.</p><p></p><p>It sounds simple, but building our software factory came down to two things. First, build a factory that produces something of quality. Then make sure you're telling it the right thing to build.</p>]]></content:encoded></item><item><title><![CDATA[Full stack means infra too]]></title><description><![CDATA[Because a product that only runs on your laptop isn&#8217;t a product]]></description><link>https://blog.cgardens.dev/p/full-stack-means-infra-too</link><guid isPermaLink="false">https://blog.cgardens.dev/p/full-stack-means-infra-too</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Tue, 13 Oct 2026 15:27:35 GMT</pubDate><content:encoded><![CDATA[<p><em><a href="https://lnkd.in/p/dA7CDUuR">Originally posted on LinkedIn.</a></em></p><p></p><p>It&#8217;s popular right now to say that with AI, every software engineer needs to be full stack. But they usually just mean frontend and backend. That&#8217;s not the full stack. The full stack is the product, the code, the infrastructure it runs on, and the dev tools you build it with. Product gets plenty of airtime. I want to talk about infra and dev tools.</p><p></p><p>Before AI, engineering specialized for good reasons. Being useful in any one layer meant years of accumulated knowledge and constant upkeep, so depth beat breadth. And the only way to ship fast was to spread the work across specialists. AI dissolved both constraints. The long tail of syntax and config is handled for you, and a single engineer with agents has the parallelism that used to require a team.</p><p></p><p>The engineer&#8217;s job is now product shipper. Not code writer, not pull request merger: the person responsible for a feature running in production. You can&#8217;t be responsible for a system you don&#8217;t understand. If your mental model ends where the app starts, you can&#8217;t ship on your own.</p><p></p><p>It&#8217;s the same story with dev tools, except faster. New models and new tools are a steady stream, and they&#8217;re step-function improvements. The most impactful engineers maintain their own tooling so they get that tech the day it ships and have it tuned to how they work. If you&#8217;re waiting for an internal tools team to deliver AI tooling to you, you&#8217;re already behind.</p><p></p><p>What&#8217;s exciting about this moment is that nobody knows what they&#8217;re doing yet. Anyone who sits down and reasons about this from the first principles is making novel discoveries. So own your tools now. Ship something end to end, including the infra. When you get stuck, use AI to get unstuck, but get through.</p><p></p><p>For years my advice to engineers has been the same: learn infra. AI has only increased the salience of that guidance.</p>]]></content:encoded></item><item><title><![CDATA[Building an AI Assistant for Enterprise SaaS]]></title><description><![CDATA[Because we all prefer to loop in 1 person instead of 10.]]></description><link>https://blog.cgardens.dev/p/building-an-ai-assistant-for-enterprise</link><guid isPermaLink="false">https://blog.cgardens.dev/p/building-an-ai-assistant-for-enterprise</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Fri, 09 Oct 2026 20:31:45 GMT</pubDate><content:encoded><![CDATA[<p><em>Originally posted on LinkedIn as a four-part series: <a href="https://lnkd.in/p/gBWGvC72">Part 1</a>, <a href="https://lnkd.in/p/g2CFRj4y">Part 2</a>, <a href="https://lnkd.in/p/eZbPCZVi">Part 3</a>, <a href="https://lnkd.in/p/gKKuZ3jT">Part 4</a>.</em></p><p>How do you build a great AI interaction for an enterprise customer who is not technical and does not want to be?</p><p></p><p>That&#8217;s what I&#8217;m working on at Levelpath right now, and I&#8217;d guess most product teams could say the same. It&#8217;s new, nobody has the patterns yet, and the ones that exist change every few months. So here&#8217;s where I&#8217;ve landed.</p><p></p><p>Underneath it there are three separate questions we&#8217;re all reasoning through:</p><ol><li><p>Do you talk to one agent or many?</p></li><li><p>How does the customer interact with agents that run on their own?</p></li><li><p>Is it one never-ending conversation, or separate sessions?</p></li></ol><p></p><h2>1. One agent or many</h2><p></p><p>The hot new pattern is a roster of named agents. Grok Bot nudges you to create a chief of staff, an inbox bot, an expenses bot, and pick who to talk to. It&#8217;s familiar. I talk to my procurement manager and my AP specialist the same way. It&#8217;s nice to know who you&#8217;re talking to.</p><p></p><p>But it pushes routing onto the customer. They have to know which agent is the right one.</p><p></p><p>We&#8217;ve been here once. When Levelpath started in 2022, LLMs could do far less and context windows were small, so we scoped the assistant to the page you were on. It frustrated users. The contracts page and the suppliers page each had an assistant, but they were different ones, and nothing you said to one carried over to the other. That&#8217;s the same problem the roster creates: nobody wants to track which agent knows what. We were already headed toward one universal assistant, the union of every role, pulling in the right skills and data as the task needs. The roster pattern only reinforced it.</p><p></p><p>Under the hood the universal assistant is still routing: to different skills, different data, effectively different specialists. There may be value in showing that. Grafana&#8217;s assistant tells you which skills it&#8217;s using as it works. You never ask for the AP specialist; it switches and you can see that it did. I think that&#8217;s useful mostly as information, and maybe more for me debugging than for the customer. Not sure yet.</p><p></p><p>In any big organization, getting one thing done means hunting down ten people. It would be simpler if one person knew it all. In human terms that&#8217;s impossible: roles exist because a person can only execute so much, so at scale you end up with a strategic sourcing manager, an AP specialist, a lawyer, a category manager. A sourcing manager agent feels familiar and safe to a company adopting AI. But it&#8217;s anthropomorphizing in a way that gives away the one thing AI is uniquely good for: a single entity that can actually do all the roles.</p><p></p><h2>2. How does the customer interact with agents that run on their own?</h2><p></p><p>A growing share of what an AI product does happens with nobody watching. An invoice arrives and an agent processes it. A supplier&#8217;s news gets scanned every day for adverse media. These agents are narrow, goal-driven, and mostly silent.</p><p></p><p>One hard problem with them is showing the customer what they did on their behalf. That&#8217;s a real design challenge, and it&#8217;s out of scope here.</p><p></p><p>What&#8217;s in scope is the moment a background agent needs a human. The adverse media agent finds something on a critical supplier. It could kick off a review process on its own, and for most suppliers that&#8217;s exactly what it does. But the customer configured it to check in before acting on anything tied to a critical supplier, so it stops and asks. Now the user has to steer something that mostly runs unattended and only occasionally surfaces.</p><p></p><p>The bot-first products handle this by making the background agent its own thing to talk to. That&#8217;s the roster problem again. The customer now has to know that the supplier risk bot exists, find it, and remember what it knows.</p><p></p><p>The answer I keep landing on is that the escalation should arrive in the same universal assistant from part 1. The agent pauses, the question shows up where the customer already is, the customer answers, the agent resumes. If the customer wants the agent to behave differently next time, they say so in that same conversation.</p><p></p><p>Under the hood I still want those agents to be discrete, with defined, readable prompts, so a non-technical admin can open one and see exactly what it will and won&#8217;t do. But that&#8217;s structure. The customer interacts with one thing.</p><p></p><p>What I haven&#8217;t settled is what that one thing is when it pauses. Is my assistant relaying a message from the supplier risk agent and carrying my answer back? Or does it take on the risk agent&#8217;s context, so I&#8217;m talking to the combination? Relay is the familiar model, a chief of staff telling you what someone on the team said. Merge is the AI-native one, since the model is the only thing in the system that can hold both contexts at once. With the right UX either can feel like a single conversation.</p><p></p><p>If you&#8217;ve built one of these, which did you pick, and would you pick it again?</p><p></p><h2>3. Is it one never-ending conversation, or separate sessions?</h2><p></p><p>The new paradigm is the conversation that never ends. Grok Bot and Muse both do this: each agent is one unbounded thread. You never start a new chat. You just keep talking to the same thing, for months.</p><p></p><p>What that buys you is the absence of a seam: no new-chat button, no deciding what carries forward. For a non-technical customer that&#8217;s appealing. They shouldn&#8217;t have to think about context at all.</p><p></p><p>The trouble starts when you ask what the customer is expecting from it. The natural expectation is that the thing remembers everything that was ever said. It doesn&#8217;t. Users report Grok Bot reloading 200 to 250k tokens on every reply, and it still summarizes the transcript as it nears the limit. It just doesn&#8217;t tell you. Muse is less forthcoming about how it works or what it costs, but the latency is similarly sluggish, and one can only guess it&#8217;s doing something close to the same thing. So the customer pays 5 to 15 times the cost of a compacted context, plus seconds on every message, for a promise of total recall that compaction quietly breaks.</p><p></p><p>There&#8217;s a plainer problem too. I want threads. (Remember Slack before threads? The horror!) On a normal day I&#8217;m working ten different things, and I want to click into each one, not scroll through a whole day of interleaved chat to find where a conversation left off. A never-ending chat might be fine for the quick back-and-forth with an assistant. For anything that&#8217;s a job, it&#8217;s the wrong shape.</p><p></p><p>What I want instead is closer to how a good human assistant works. They haven&#8217;t memorized everything you&#8217;ve said to them over the years. What they have is a sense of how you like to work and what you do often. Hand them a new job and they know whether it&#8217;s a repeat, and if it is, they can do most of it or at least tell you what comes next.</p><p></p><p>That&#8217;s a fine starting model. New thing to do, hit new. That&#8217;s the whole burden on the user. Everything about that job stays in its thread until the job is done, whether that takes a minute or a month. The assistant brings how you work and what you&#8217;ve done before. Things you ask for repeatedly become skills. If this is something you did last quarter, it says so and suggests where to start, instead of pretending it never forgot. And where it beats the human assistant: it has every past transcript and all the data, and it can look any of it up in seconds. The exact record comes from a lookup, not from hoping it&#8217;s still in the context.</p><p></p><p>If you&#8217;ve lived with the never-ending version for a while, how is it working for you?</p><h2>4. What it costs to build</h2><p></p><p>Across the first three sections the answer kept coming out the same way. Named agents, a separate bot for each background job, one thread that never ends: each is a real piece of the system&#8217;s structure pushed out onto the customer. Every time I checked a design against how I&#8217;d treat a human in that role, the human analogy helped a little and then started costing something only the AI could give. So the spec is short. One assistant, every role at once, new thread per job, memory for the shape of how you work, lookup for anything exact.</p><p></p><p>That is a single entity that is simultaneously the sourcing manager, the AP specialist, the category manager, and the lawyer, that knows how you work without being told, that has read every conversation you&#8217;ve had with it and can find any one of them in seconds, and that runs a dozen jobs for you in parallel without you keeping track of any of them. No human has ever been that. That&#8217;s what we&#8217;re asking for, and the only reason it&#8217;s askable is that the intelligence behind it is not a person.</p><p></p><p>None of that structure goes away. It moves to engineering. Here is what we have to solve to make the customer&#8217;s version that simple.</p><p></p><p>Routing. The customer no longer picks the specialist, so the assistant has to, on every turn: which skills, which data, and only the permissions this job calls for. Too loose and it answers an AP question like a sourcing manager. Too tight and it can&#8217;t do the job. Too broad and it has access it shouldn&#8217;t.</p><p></p><p>Context assembly. Every new thread starts empty. Filling it with the right working patterns, skills, data, and prior runs, inside a budget that keeps it fast, is the whole game.</p><p></p><p>Memory. How you work and what you do often should come forward on its own. What was actually said or done should never sit in context by default. It comes from fast, permission-aware retrieval that the assistant trusts more than its own guess.</p><p></p><p>That&#8217;s my current thesis on the ideal assistant inside an enterprise SaaS product. I haven&#8217;t seen anyone fully build it yet, and some of what it needs are hard engineering problems I haven&#8217;t seen anyone else solve either. Those are the problems we&#8217;re working on now.<br></p>]]></content:encoded></item><item><title><![CDATA[How Wordle Taught Me to Use AI]]></title><description><![CDATA[Because the WordleBot and I are not playing the same game, and neither is the LLM]]></description><link>https://blog.cgardens.dev/p/how-wordle-taught-me-to-use-ai</link><guid isPermaLink="false">https://blog.cgardens.dev/p/how-wordle-taught-me-to-use-ai</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Mon, 05 Oct 2026 15:35:39 GMT</pubDate><content:encoded><![CDATA[<p><em><a href="https://www.linkedin.com/feed/update/urn:li:activity:7512874470739697664/">Originally posted on LinkedIn.</a></em></p><p></p><p>The NYT WordleBot's explanations of my Wordle games are often completely foreign to me. The bot knows every possible answer and treats each turn as a search problem, the kind usually modeled with Shannon entropy. It picks the guess that minimizes the expected number of turns left, which often means playing a word that cannot be the answer because it splits the remaining pool more cleanly. Even<a href="https://www.nytimes.com/2022/04/07/upshot/wordle-bot-introduction.html"> its creators note</a> that the best guess for a computer is not necessarily the best guess for you.</p><p></p><p>I play by pushing letters around until something jogs my memory. The skeleton is the same, prune first and then commit, but my version is shaped around the tools I have. I do better once a couple of consonants are pinned, because consonants are what trigger my recall. The bot and I are not disagreeing. We are solving the same problem with different technology.</p><p></p><p>My only real shot at beating the bot on any given day is a high-variance move: commit to a real guess early and hope it lands. That gets the occasional win and is even worse in the long run, though I do not track my scores, so I will never find out how much worse. Over any real number of games the bot wins, because Wordle is shaped exactly like the problems deterministic computing is built for: a closed search space with a clear way to score every guess. When a problem fits the machine that well, it is no contest.</p><p></p><p>Same lesson, new machine</p><p></p><p>An LLM has different technology at its disposal. It reads about three orders of magnitude faster than I do and writes about two orders of magnitude faster. It writes code and operates a computer fluently, so anything that can be done through a terminal it can do at machine speed. It does not get bored, so it will grind through repetitive work I would quietly cut corners on. It is also sycophantic and bad at checking its own work.</p><p></p><p>So my job is to reshape problems around that profile: point the model at work that plays to its strengths, and build a harness that checks its output where it is weak.</p><p></p><p>Do that and the execution capacity is effectively unlimited. Two shapes keep coming up. If a problem can be turned into code, a model is great at it. If it can be turned into search, with a clear way to test candidates, a model can grind its way to an answer. In September, OpenAI <a href="https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/">announced a resolution </a>of the Navier-Stokes Millennium Prize problem produced by roughly 10,000 agents running for 88 hours. Benn Stancil calls this <a href="https://benn.substack.com/p/brute-intelligence">brute intelligence: n</a>ot monkeys at typewriters, but a million researchers taking swings at the same problem. No human mathematician would approach it that way, the same way no human plays Wordle like the bot.</p>]]></content:encoded></item><item><title><![CDATA[We plugged bugs into the software factory]]></title><description><![CDATA[Because when the bug report is clear, the human is just latency]]></description><link>https://blog.cgardens.dev/p/we-plugged-bugs-into-the-software</link><guid isPermaLink="false">https://blog.cgardens.dev/p/we-plugged-bugs-into-the-software</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Sat, 03 Oct 2026 04:02:11 GMT</pubDate><content:encoded><![CDATA[<p><em><a href="https://www.linkedin.com/feed/update/urn:li:activity:7495479748656107520/">Originally posted on LinkedIn.</a></em></p><p>We plugged bugs into the software factory. Every night the agentic loop works through open bug tickets, fixes them, and opens PRs. The first thing engineers do in the morning is go through the queue. Most go through without comment.</p><p></p><p>The non-obvious part is that this only works if ticket quality is high. The number one reason a bug can't be fixed is that it's vague about what the expected behavior should be. That isn't an agent problem. Those are the same tickets that would have taken rounds of back and forth with a human developer. We automated the ticket quality check, so now the gap gets caught at intake instead of after someone picks it up.</p><p></p><p>I wrote before about each major area being owned by one or two engineers. In the pre-AI world that headcount plus this bug load meant nothing below P1 ever got touched. Like most teams, we had a standing backlog of P2s and P3s nobody was realistically going to reach. Now P0 through P3 are all handled within SLA, and the whole review takes about as long as a cup of coffee.</p><p></p><p>One casualty: we used to celebrate a bug fix champion of the week. The auto fixer took it the first week we turned it on and hasn't given it back since. Three weeks running, no human close.</p><p></p><p>The real point is that bugs were a plug-in. Once you have a factory that turns Jira tickets into high-quality features, you can point it at a lot of different kinds of work. That's the leverage: not any one automation, but everything that becomes possible once the machine exists.</p>]]></content:encoded></item><item><title><![CDATA[We replaced squads with single owners]]></title><description><![CDATA[Because one mind was always the ideal]]></description><link>https://blog.cgardens.dev/p/we-replaced-squads-with-single-owners</link><guid isPermaLink="false">https://blog.cgardens.dev/p/we-replaced-squads-with-single-owners</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Mon, 17 Aug 2026 21:37:01 GMT</pubDate><content:encoded><![CDATA[<p><em>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_a-squad-of-eight-on-a-product-area-usually-activity-7495230967872217088-L4kH?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</em></p><p><span>A squad of eight on a product area usually had one or two people who actually understood the system. The rest were following. They wrote code against decisions someone else had made, and the depth lived in a couple of heads. When someone wasn't pulling their weight, the team absorbed it and nobody could see it clearly.<br><br>In the AI age we staff one or two engineers per area.<br><br>This isn't a headcount story. It's the opposite: the same engineers now cover far more surface area, and we can sprint on more of the product at once than we ever could with squads.<br><br>The work that made large teams necessary moved. Execution capacity comes from the LLM. Most bug work runs in the background. The engineer's job moved up a level: build the system that writes the code, encode enough of the product intent that what comes out is right, and keep the whole thing running as the product underneath it changes. That's a harder job than writing the feature yourself. You're not solving one problem, you're designing the conditions under which a whole class of problems gets solved correctly without you.<br><br>One person owning an entire area used to be impossible because bug fixing alone would consume them. Keep up with the bugs and no new features ever get built. Now that volume runs in the background, and what remains is the part that never parallelized well anyway: knowing what should be built, and whether what shipped is right.<br><br>The second-order effects matter more than the speed. No handoff between the person who understood the requirement and the person who wrote the code. No context lost across a standup. And output per engineer is finally legible, because there's one name on the area.<br><br>This is why we're building an owners-only culture. The job isn't writing code anymore. It's owning a machine that writes code, and owning what it ships.</span></p>]]></content:encoded></item><item><title><![CDATA[AI is your engineering execution capacity now]]></title><description><![CDATA[Because your engineering capacity is only bound by how many tokens you are willing to spend]]></description><link>https://blog.cgardens.dev/p/ai-is-your-engineering-execution</link><guid isPermaLink="false">https://blog.cgardens.dev/p/ai-is-your-engineering-execution</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Mon, 17 Aug 2026 21:11:41 GMT</pubDate><content:encoded><![CDATA[<p>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_ai-is-a-limitless-supply-of-engineering-capacity-activity-7486117331312484352-9U0Q?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</p><p><span>AI is a limitless supply of engineering capacity. Most teams can't use it. Our agentic software factory let us harness that capacity to rebuild our entire product in a quarter.<br><br>The bottleneck for most teams older than a year is the development environment AI has to operate in. A few months ago, AI agents couldn't even reliably start our application. The best any engineer could do was work one local task at a time and manually hand sub-tasks to a model. All that capacity, and no way to plug into it.<br><br>So we rebuilt from the ground up. The conceit from day one: the entire app would be built by the software factory. That was the forcing function. Every part of the development environment had to work for agents, because agents were doing all the work. No retrofitting agentic development onto the codebase. We expected a tradeoff, going slow now to go fast later. There wasn't one. It was faster from the start.<br><br>The result: zero lines of code to live customers in one quarter. Four years of product rebuilt, plus new features, in that same window. Engineers who used to run one local task now oversee 5+ in parallel. Conservatively, a 5x productivity increase, and we haven't found the ceiling.<br><br>The AI world is all about speed. Getting it means working backwards from a single question: how do we harness this capacity? Everything about your engineering system follows from the answer. Get it right and you're running on fusion while everyone else is still running on steam.</span></p>]]></content:encoded></item><item><title><![CDATA[Managing context for complex policy]]></title><description><![CDATA[Because with the right harness you make your LLM follow the rules]]></description><link>https://blog.cgardens.dev/p/managing-context-for-complex-policy</link><guid isPermaLink="false">https://blog.cgardens.dev/p/managing-context-for-complex-policy</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Mon, 17 Aug 2026 21:10:17 GMT</pubDate><content:encoded><![CDATA[<p><em>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_how-routing-trees-solve-the-enterprise-ai-activity-7483216563576156160-hfQD?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</em></p><p><span>The hard part of enterprise agents isn't knowledge. It's timing. The model doesn't just need to know the customer's policies, it needs to apply the right one at the right step of an orchestrated workflow. Stuffing all 400K tokens of rules into the prompt gave our agent perfect knowledge and terrible timing: hallucinated contacts, skipped steps, a lost thread.<br><br>The industry has an answer for this, sort of. Route the model to the context it needs instead of stuffing the prompt. But the public playbook stops at one hop: classify the intent, load the matching instructions, done. Salesforce topics work this way. Skills and progressive disclosure work this way.<br><br>One hop can't fix a timing problem. The decisions are chained, and each link needs its own rules. <br><br>Past that first hop, we were on our own. No write-up existed for what we needed, so we built it. A routing tree: 72 nodes, each owning a single decision, each carrying only the documents it needs. The agent traverses one decision at a time and never sees more than 5-15K tokens per turn. Token usage dropped 95%. The hallucinations stopped.<br><br></span><strong><a href="https://www.linkedin.com/in/nikolaupite/"><span>Nikola Up&#299;te</span></a></strong><span> and </span><strong><a href="https://www.linkedin.com/in/anna-bicevska/"><span>Anna Bicevska</span></a></strong><span> developed node-based routing to handle keeping agents on track for navigating the most complicated enterprise policies in chat. <br><br>Full writeup: </span><strong><a href="https://lnkd.in/g3258Dpy">https://lnkd.in/g3258Dpy</a></strong></p>]]></content:encoded></item><item><title><![CDATA[Specialized models are viable for real engineering work]]></title><description><![CDATA[Because you don't need the frontier model for everything]]></description><link>https://blog.cgardens.dev/p/specialized-models-are-viable-for</link><guid isPermaLink="false">https://blog.cgardens.dev/p/specialized-models-are-viable-for</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Mon, 17 Aug 2026 21:07:44 GMT</pubDate><content:encoded><![CDATA[<p><em>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_specialized-models-are-now-viable-for-real-activity-7481005597367148544-aSy7?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</em></p><p><span>Specialized models are now viable for real engineering work. Almost all of our development runs through an autonomous loop, and we'd been on a trajectory where each new frontier model made that loop markedly better. With the latest generation, that's no longer the pattern.<br><br>We run that loop on Cursor Cloud Agents. We started there because the harness was easy to stand up, and it's the most ergonomic remote agent environment we've used. We assumed the coding in that loop needed the best general model available.<br><br>We tried Composer on a lark. We were blown away: for most of the coding we run through the loop, it was a drop-in replacement. It's faster too. We haven't benchmarked it precisely, but the gap is obvious, and public numbers put it near 50% quicker on coding tasks. It's also 90% cheaper than Opus.<br><br>It isn't good at everything. Behind our QA agent it was noticeably worse. That's the point. It's specialized for coding.<br><br>For us the tradeoff flipped. A cheaper, faster, good-enough coding model beats a general model that's stronger at things the task never touches. Composer is the good-enough one today. I'd expect that to flip: specialized coding models pulling ahead of general ones at coding outright, and getting cheaper and faster as they do.<br><br>QA is the gap for us now, so if you know a model specialized for it, tell me.<br>I'm surprised I don't hear more about Composer. And it raises a real question for model specialization: who wins there? Maybe it's the major foundation labs that already hold all the leverage. Maybe specialization is the crack where someone else gets in.</span></p>]]></content:encoded></item><item><title><![CDATA[AI is more qualified to approve PRs than you are]]></title><description><![CDATA[Because humans were already rubbish at code review]]></description><link>https://blog.cgardens.dev/p/ai-is-more-qualified-to-approve-prs</link><guid isPermaLink="false">https://blog.cgardens.dev/p/ai-is-more-qualified-to-approve-prs</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Mon, 17 Aug 2026 21:06:02 GMT</pubDate><content:encoded><![CDATA[<p>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_theres-a-lot-of-handwringing-about-merging-activity-7480652002952146944-Ikvb?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</p><p><span>There's a lot of handwringing about merging code that has only been reviewed by AI. It's backwards. A process where some code is reviewed only by AI will be more security conscious than what we do today.<br><br>Start with the problem. AI writes more code than any team can review, so everything gets the same rushed skim, and the changes that actually matter don't get the attention they deserve. The fix isn't more reviewers or going slower. Teams can&#8217;t afford to ship slower in this environment. It's deciding what humans need to see at all.<br><br>Humans were always bad at this part anyway. Outside the most egregious cases, we've never been good at spotting a bug or a vulnerability by reading code cold. AI has been better at that for a long time. A half-distracted developer waving through a diff is not your security baseline. An agent making that call catches more.<br><br>So route deliberately. Certain file paths and high-risk changes go to humans. That's a mix of deterministic rules and AI classification, and classification is something AI is good at. Everything else is reviewed by AI alone. Now the sensitive code gets real scrutiny instead of competing for attention with everything else.<br><br>What was human review ever really verifying? Legibility. Could another person read this and follow it. But the primary reader is shifting from human to AI, so stop optimizing for human legibility and start optimizing for AI legibility. We used to have complexity metrics for how hard code was for a person to parse. I want the equivalent for a model, scoring not just structure but whether the prompts, skills, and agents around the code help it understand what it's looking at. That's the thing worth building.</span></p>]]></content:encoded></item><item><title><![CDATA[The AI-Pilled 4 Yorkshiremen (Tech Exec Edition)]]></title><description><![CDATA[Because the AI era requires some good-humored fatalism]]></description><link>https://blog.cgardens.dev/p/the-ai-pilled-4-yorkshiremen-tech</link><guid isPermaLink="false">https://blog.cgardens.dev/p/the-ai-pilled-4-yorkshiremen-tech</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Mon, 17 Aug 2026 21:04:33 GMT</pubDate><content:encoded><![CDATA[<p><em>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_the-ai-pilled-4-yorkshiremen-tech-exec-edition-activity-7475552072721973248-F4YY?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</em></p><p><span>"Our engineers use Cursor."<br><br>"Use Cursor? We moved beyond IDEs entirely. Engineers just describe features to agents."<br><br>"Describe features? That's adorable. Our PMs describe features directly to agents."<br><br>"PMs? We eliminated PMs. Customers describe features directly to agents."<br><br>"Customers? We eliminated customers. Agents analyze user behavior and create the roadmap themselves."<br><br>"Roadmap? Our agents rewrite the company strategy every night."<br><br>"Strategy? We don't even have executives anymore. The agents found them to be a bottleneck."<br><br>"Executives? Luxury. Our board meetings are attended exclusively by agents."<br><br>"Board meetings? We haven't had a board meeting in months. The agents acquired three companies while we were asleep."<br><br>"Companies? We don't have companies. Just a loose federation of cooperating agents with a Delaware LLC."<br><br>-- Inspired by Monty Python's 4 Yorkshiremen sketch: </span><strong><a href="https://lnkd.in/dA9U4b9E">https://lnkd.in/dA9U4b9E</a></strong></p>]]></content:encoded></item><item><title><![CDATA[The AI can do it]]></title><description><![CDATA[Because we are conditioned to assume that we are special]]></description><link>https://blog.cgardens.dev/p/the-ai-can-do-it</link><guid isPermaLink="false">https://blog.cgardens.dev/p/the-ai-can-do-it</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Mon, 17 Aug 2026 21:03:36 GMT</pubDate><content:encoded><![CDATA[<p><em>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_you-cant-build-an-ai-native-product-with-activity-7475189679043141633-RSZf?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</em></p><p><span>You can't build an AI-native product with a team that isn't AI-native.<br><br>Most teams still think about AI the wrong way. They start with a human process and ask: "How can AI help people do this faster?" That's the wrong question.<br><br>An AI-native team starts with the outcome and asks: "How do we get the AI to achieve this outcome?" Then they work backwards from there. Maybe the AI needs verification. Maybe it needs an approval threshold. Maybe it needs an audit trail. Maybe there are edge cases that still require human judgment.<br><br>The software becomes scaffolding around the AI, not the other way around.<br><br>This is a hard mindset shift because for most of business history, outcomes were constrained by people. If you wanted more contract reviews, more supplier research, more invoice matching, or more policy enforcement, you hired more people. The software was there to support the human.<br><br>Now, for a growing set of tasks, you can spend tokens instead. That doesn't mean humans disappear. It doesn't mean AI can do everything. But it does mean we should stop assuming humans are the default actor in every business process.<br><br>Procurement is full of work that AI can already do as well as, and often better than, people: three-way invoice matching, sending NDAs, supplier research, tracking RFP responses, policy enforcement.<br><br>The opportunity isn't to help humans do those tasks faster. The opportunity is to have AI do them, then add the controls, verification, and auditability required to trust the result.<br><br>The most important mindset shift in AI is moving from: "How do we use AI in this process?" to "Why does a human need to do this at all?"</span></p>]]></content:encoded></item><item><title><![CDATA[Integrations in the Age of AI]]></title><description><![CDATA[Because the AI is happy to make up what it can't find]]></description><link>https://blog.cgardens.dev/p/integrations-in-the-age-of-ai</link><guid isPermaLink="false">https://blog.cgardens.dev/p/integrations-in-the-age-of-ai</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Mon, 17 Aug 2026 21:02:00 GMT</pubDate><content:encoded><![CDATA[<p>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_the-bottleneck-to-scaling-enterprise-ai-is-activity-7460682834144477184-RBG0?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</p><p><span>The bottleneck to scaling enterprise AI is almost never the model. It's the data underneath it. I've spent most of my career in data infrastructure. That bias shapes how we build at Levelpath.<br><br>The customers we work with who are building agents run into the same wall: the agent can't access the right data to make a decision. A supplier record that lives in a system it can't reach. An approval history that was never structured. A risk signal that exists somewhere, just not somewhere the agent can see. When that happens, agents don't tell you they're stuck. They fill in the gap and keep going. That's the failure mode nobody talks about enough.<br><br>At Levelpath, we do the integration work upfront. Every relevant data source connected, normalized, accessible. We do this because our approach is agent-first: everything in the product is designed to be done by an agent by default. For that to work, the agent needs to be trustworthy. It escalates when it should and it never makes something up. We build those guarantees in.<br><br>Applied AI is usually a data and integration problem. The models are plenty capable. Give them the right data and they'll do the work.</span></p>]]></content:encoded></item><item><title><![CDATA[AI-Managed Infrastructure]]></title><description><![CDATA[Because Terraform was never meant for humans]]></description><link>https://blog.cgardens.dev/p/ai-managed-infrastructure</link><guid isPermaLink="false">https://blog.cgardens.dev/p/ai-managed-infrastructure</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Wed, 13 May 2026 03:02:17 GMT</pubDate><content:encoded><![CDATA[<p><em>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_given-the-claude-deleted-a-production-database-share-7455486085788516353-dXhs?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</em></p><p>Given the <a href="https://x.com/lifeof_jer/status/2048103471019434248">Claude-deleted-a-production-database</a> story going around, now seems like a good time to advocate for AI managing your infrastructure.</p><p>Last week I granted AdministratorAccess to Claude Code for an isolated AWS account. Asked it to set up Grafana. 17 minutes. Then the full LGTM observability stack including supporting infrastructure (ECS, RDS, volumes, networking, etc) and a working hello world app to prove that the stack was working. 22 minutes.</p><p>I&#8217;d been meaning to properly explore Grafana for years. Infra setup was always one of the hurdles that made it feel like a big undertaking. That blocker is gone now.</p><p>No, don&#8217;t give Claude access to your production environment. But it is time to identify what environments we can hand off to AI<em>.</em></p><p>A lot of what slows down software innovation is skill cleavages &#8212; the handoffs between frontend and backend, between application engineers and DevOps. Every boundary between specializations is a place where one person&#8217;s idea has to become someone else&#8217;s ticket.</p><p>Those boundaries are collapsing. I no longer need to go through a DevOps team to spin up infrastructure. The side app we built to track our token spend just runs. Claude manages the infra. If it deletes the database, we&#8217;re back in five minutes.</p><p>This means I can run software that previously I simply wouldn&#8217;t have. Not because I couldn&#8217;t write it, but because owning the full stack was too much overhead for a side project or an experiment.</p><p>So yes, be careful with production. But I&#8217;ll be disappointed if AI doesn&#8217;t delete at least one of our non-prod environments this year. If it doesn&#8217;t, it probably means we&#8217;re leaving velocity and innovation on the table.</p>]]></content:encoded></item><item><title><![CDATA[CI in Agentic Times]]></title><description><![CDATA[Because agents actually follow up on clean up work]]></description><link>https://blog.cgardens.dev/p/ci-in-agentic-times</link><guid isPermaLink="false">https://blog.cgardens.dev/p/ci-in-agentic-times</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Wed, 29 Apr 2026 18:00:57 GMT</pubDate><content:encoded><![CDATA[<p><em>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_agentic-loops-change-what-belongs-in-ci-activity-7454899738967166976-u4ew?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</em></p><p>Agentic loops change what belongs in CI. Previously you were trading off delivery speed against codebase hygiene.</p><p>Block on linting, coverage, pattern enforcement, you slow the team. Don&#8217;t block, the codebase rots.</p><p>Fast static analysis and formatting checks still belong at the gate. The shift is with the slow, expensive stuff. Running full coverage analysis, enforcing architectural patterns, refactoring for DRY used to be a real tradeoff. Post-merge meant a context switch back to something a developer had already moved on from. That work was an expensive context switch for a developer and usually got skipped.</p><p>That cost doesn&#8217;t exist when an agent is doing the work. At Levelpath we&#8217;ve already moved coverage tracking and developer experience metrics to background agents running post-merge. We&#8217;re now looking at what else moves: pattern enforcement, refactoring, architectural consistency.</p><p>The bar for codebase hygiene stops being what&#8217;s worth slowing people down and starts being what standard you actually want to hold.</p>]]></content:encoded></item><item><title><![CDATA[AI is a new substrate for Algorithms]]></title><description><![CDATA[Because we all wanted a new edition of Cracking the Coding Interview]]></description><link>https://blog.cgardens.dev/p/ai-is-a-new-substrate-for-algorithms</link><guid isPermaLink="false">https://blog.cgardens.dev/p/ai-is-a-new-substrate-for-algorithms</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Tue, 28 Apr 2026 17:54:52 GMT</pubDate><content:encoded><![CDATA[<p><em>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_for-decades-good-algorithms-were-shaped-share-7452072630293094400-XPwK?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</em></p><p>For decades, good algorithms were shaped by the silicon running them. Cache lines, branch prediction, pipeline stalls &#8212; the hardware&#8217;s characteristics shaped the code.</p><p>AI algorithms deserve the same mindset.</p><p>LLMs have performance characteristics too. Fast and tireless on some things, unreliable on others. Better at generation than verification, or sometimes the reverse. Context has a sweet spot. Token costs and latency stack fast. The algorithms we build on top of models should be shaped by all of this.</p><p>When I think about an AI system at Levelpath, it mostly comes down to two decisions: how to decompose a problem, and how to structure the creator/evaluator loop around each piece. Decompose too coarsely, the model drowns. Too finely, orchestration overhead eats the gains. Strong creator, weak evaluator: you ship garbage confidently. Flip it: you burn tokens forever.</p><p>To make it more challenging, these characteristics keep moving. Context windows grow. Reasoning gets cheaper. What was a 10-step decomposition last year is one prompt today. We design for current models, knowing that we can pull away constraints as the substrate evolves.</p><p>The software engineer&#8217;s job is changing, and the shape of the new job is this: designing algorithms for the AI substrate.</p>]]></content:encoded></item><item><title><![CDATA[Self-Healing DX Pipelines]]></title><description><![CDATA[Because we can afford to measure new things if our metrics pipelines maintain themselves]]></description><link>https://blog.cgardens.dev/p/self-healing-dx-pipelines</link><guid isPermaLink="false">https://blog.cgardens.dev/p/self-healing-dx-pipelines</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Mon, 27 Apr 2026 17:51:44 GMT</pubDate><content:encoded><![CDATA[<p><em>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_we-track-dx-metrics-via-a-github-action-at-share-7452085017180102656-tK2D?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</em></p><p>We track DX metrics via a GitHub Action &#8212; code coverage, CI build time, app start time, hot reload time.</p><p>We&#8217;re mid-replatform, so the app and how it runs are changing constantly. What we want to measure isn&#8217;t. Result: the action breaks every other morning. A path moves, a script is renamed, a dependency shifts.</p><p>For a week I just fixed it: open the failure, paste it into Claude Code, patch the workflow, merge. I was a courier between the CI log and the model.</p><p>So I built a second Action: headless Claude Code, with read access to the failing run&#8217;s logs and permission to open a PR against the workflow. When the DX metrics action fails, it reads why, proposes the fix, opens the PR. I review and merge.</p><p>Most AI energy right now goes into building new agentic workflows. The quieter move is pointing agents at the workflows you already have &#8212; to keep them doing the right thing as everything underneath shifts.</p><p>When the <em>what</em> is stable (measure these four DX metrics) and only the <em>how</em> keeps changing, an LLM is enough to hold it together. Self-healing is a declarative way of describing what you want. The model handles the imperative.</p>]]></content:encoded></item><item><title><![CDATA[Inverting the Developer and AI Relationship]]></title><description><![CDATA[Because there are more tokens than human-hours in the day]]></description><link>https://blog.cgardens.dev/p/inverting-the-developer-and-ai-relationship</link><guid isPermaLink="false">https://blog.cgardens.dev/p/inverting-the-developer-and-ai-relationship</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Mon, 20 Apr 2026 20:11:37 GMT</pubDate><content:encoded><![CDATA[<p><em>Originally posted on <a href="https://www.linkedin.com/posts/cgardens_when-a-jira-ticket-is-created-on-our-team-share-7452061518868799489-9VYc?utm_source=share&amp;utm_medium=member_desktop&amp;rcm=ACoAAAgTj18B0dc65XxH74VWlI5_UX1E1BVt6lE">LinkedIn</a>.</em></p><p>When a ticket is created on our team, an AI agent picks it up by default. A human only gets pulled in if the agent escalates.</p><p>How did we get here? In the beginning, there were engineers. Then engineers got AI assistants, and the assistant sat next to the human: you&#8217;d pick up a ticket, figure out what was worth delegating, and hand pieces of it to the model. The human was the one actually working the ticket. AI was optimizing the developer&#8217;s throughput.</p><p>The key shift that we made 3 weeks ago was that by default AI does the work. Engineers spend their time unblocking them: giving them context, shaping the environment, deciding what should trigger an escalation, fixing the cases where they get stuck. Our biggest bottleneck now is code review.</p><p>The job of the engineer stops being &#8220;do tickets&#8221; and starts being &#8220;design the conditions under which tickets get done.&#8221;</p><p>The results so far: 5 engineers running this system ship roughly as much as the other 50 on our engineering team combined.</p><p>And this is V1. It&#8217;s the dumbest the system will ever be. V2 is already underway. We are building a machine that turns tokens into product delivery capacity. Now the developer is optimizing the AI&#8217;s throughput.</p><p>If you&#8217;re trying to get more out of AI in engineering, the question probably isn&#8217;t &#8220;how do I delegate better.&#8221; It&#8217;s &#8220;what would it take for the agent to pick up the work by default.&#8221;</p>]]></content:encoded></item><item><title><![CDATA[Of Course This Blog is Written by AI]]></title><description><![CDATA[Because &#8220;never published&#8221; to &#8220;done in 15 minutes&#8221; is an infinity percent improvement]]></description><link>https://blog.cgardens.dev/p/of-course-this-blog-is-written-by</link><guid isPermaLink="false">https://blog.cgardens.dev/p/of-course-this-blog-is-written-by</guid><dc:creator><![CDATA[charles]]></dc:creator><pubDate>Sat, 24 Jan 2026 15:01:07 GMT</pubDate><enclosure url="https://substackcdn.com/image/fetch/$s_!E_4-!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<p><strong>I am not a Writer. I am a tech exec, a dad, a software engineer, and a human with ideas I want to share.</strong> AI has been a huge unlock for me in doing this. Pre-AI, I struggled to publish anything. While I had plenty of ideas that were baked enough to share, the actual process of translating them from my head into written text was very slow. This struggle is evidenced by this<a href="https://blog.cgardens.dev/p/tool-for-the-job"> early post</a> on this Substack, where all I managed to do was write about how I wasn&#8217;t writing anything.</p><p>My process for writing articles is straightforward: I start with a topic I want to discuss. I then build a fairly dense outline, complete with examples and a point-by-point breakdown of what I&#8217;m trying to say. <strong>I know what I want to say.</strong> I use AI to help augment my research along the way. When it&#8217;s time to write, I pass the outline to a personal GPT, which has been trained on my past, pre-AI content to capture something of my voice. It then generates the text. There is an iteration process there where we work together to refine it. I also ask other LLMs to challenge the work to pressure test both the ideas and the writing.</p><p>Does this diminish the content of this blog? I don&#8217;t think so.</p><p>My brother texted me the other day, saying, &#8220;Some of your blog sounds like it was written by AI.&#8221; I responded, &#8220;Of course it does, because it is!&#8221; I understand his reaction, because I do it too: when I consume something and begin to suspect it was AI-written, my hackles go up, and I immediately become more skeptical of its quality. I believe we need to move past that reaction. This is related to an <a href="https://blog.cgardens.dev/p/definitive-guide-to-ai-slop">earlier piece</a> where I built a mental model for how to grapple with the use of AI in media content.</p><div class="captioned-image-container"><figure><a class="image-link image2 is-viewable-img" target="_blank" href="https://substackcdn.com/image/fetch/$s_!E_4-!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png" data-component-name="Image2ToDOM"><div class="image2-inset"><picture><source type="image/webp" srcset="https://substackcdn.com/image/fetch/$s_!E_4-!,w_424,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png 424w, https://substackcdn.com/image/fetch/$s_!E_4-!,w_848,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png 848w, https://substackcdn.com/image/fetch/$s_!E_4-!,w_1272,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png 1272w, https://substackcdn.com/image/fetch/$s_!E_4-!,w_1456,c_limit,f_webp,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png 1456w" sizes="100vw"><img src="https://substackcdn.com/image/fetch/$s_!E_4-!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png" width="514" height="472" data-attrs="{&quot;src&quot;:&quot;https://substack-post-media.s3.amazonaws.com/public/images/57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png&quot;,&quot;srcNoWatermark&quot;:null,&quot;fullscreen&quot;:null,&quot;imageSize&quot;:null,&quot;height&quot;:472,&quot;width&quot;:514,&quot;resizeWidth&quot;:null,&quot;bytes&quot;:24099,&quot;alt&quot;:null,&quot;title&quot;:null,&quot;type&quot;:&quot;image/png&quot;,&quot;href&quot;:null,&quot;belowTheFold&quot;:false,&quot;topImage&quot;:true,&quot;internalRedirect&quot;:&quot;https://blog.cgardens.dev/i/185579641?img=https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png&quot;,&quot;isProcessing&quot;:false,&quot;align&quot;:null,&quot;offset&quot;:false}" class="sizing-normal" alt="" srcset="https://substackcdn.com/image/fetch/$s_!E_4-!,w_424,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png 424w, https://substackcdn.com/image/fetch/$s_!E_4-!,w_848,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png 848w, https://substackcdn.com/image/fetch/$s_!E_4-!,w_1272,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png 1272w, https://substackcdn.com/image/fetch/$s_!E_4-!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F57444edb-5f28-4f8a-a9d3-b56719d4449b_514x472.png 1456w" sizes="100vw" fetchpriority="high"></picture><div class="image-link-expand"><div class="pencraft pc-display-flex pc-gap-8 pc-reset"><button class="pencraft pc-reset pencraft icon-container restack-image buttonBase-GK1x3M" tabindex="0" type="button"><svg aria-hidden="true" width="20" height="20" viewBox="0 0 20 20" fill="none" stroke-width="1.5" stroke="var(--color-fg-primary)" stroke-linecap="round" stroke-linejoin="round" xmlns="http://www.w3.org/2000/svg" class="icon-noB79L"><g><path d="M2.53001 7.81595C3.49179 4.73911 6.43281 2.5 9.91173 2.5C13.1684 2.5 15.9537 4.46214 17.0852 7.23684L17.6179 8.67647M17.6179 8.67647L18.5002 4.26471M17.6179 8.67647L13.6473 6.91176M17.4995 12.1841C16.5378 15.2609 13.5967 17.5 10.1178 17.5C6.86118 17.5 4.07589 15.5379 2.94432 12.7632L2.41165 11.3235M2.41165 11.3235L1.5293 15.7353M2.41165 11.3235L6.38224 13.0882"></path></g></svg></button><button class="pencraft pc-reset pencraft icon-container view-image buttonBase-GK1x3M" tabindex="0" type="button"><svg xmlns="http://www.w3.org/2000/svg" width="20" height="20" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round" class="lucide lucide-maximize2 lucide-maximize-2 icon-noB79L"><polyline points="15 3 21 3 21 9"></polyline><polyline points="9 21 3 21 3 15"></polyline><line x1="21" x2="14" y1="3" y2="10"></line><line x1="3" x2="10" y1="21" y2="14"></line></svg></button></div></div></div></a></figure></div><p>Similar to that article, I think we can build a truth table. One axis is whether the ideas being described are AI-Generated or Human-Generated. For where AI is now, none of the ideas it generates are interesting to read. It truly feels like slop. Now, of course, a lot of human-generated ideas are not that interesting either. But for the sake of our mental model, we&#8217;re going to assume a decent hit rate on humans sharing interesting thoughts and wrestling with hard problems.</p><p>The second axis is how the idea gets presented. As a software engineer, I think of the presentation as merely a &#8220;display layer.&#8221; In engineering, we&#8217;d consider whether functionality was being exposed via an API, a GUI, a mobile app, a chatbot, etc. These are different display layers. Genuinely useful products can become useless because of a bad interface. However, a fundamentally unuseful product cannot be saved by a nice display layer.</p><p>Given the time I am willing to put in, I cannot create a high enough quality display layer for my work. Yes, I could probably spend years learning the craft, but that opportunity cost is too high. AI lowers the barrier, enabling me to produce readable content quickly. The other axis on our truth table, therefore, is whether the &#8220;display layer,&#8221; or the text itself, is generated. We should be totally okay with the quadrant where the idea is human-generated, but the text is AI-written. That certainly doesn&#8217;t diminish the fully human-generated and written content. We can safely ignore all AI-generated ideas (at least for now).</p><p>My commitment in writing this blog is that I will never publish AI-generated ideas. I will not say, &#8220;Hey Claude, write an article in my voice about unit tests.&#8221; I will continue to take a human-generated thesis or outline and ask Claude to turn it into text. We are going to live in the top-right corner of that truth table.</p><p>I am not a <strong>W</strong>riter. My differentiation is not in the craft or style of my prose; the prose is simply a means of providing a display layer for the conversations and theses I want to share. With AI, I can take an idea that I&#8217;ve already baked and produce a full article in 15 minutes at a higher level of polish than if I put in hours. For people who <em>are</em> Writers, I understand why relying on AI to write the final text generally would not make sense; the prose itself is part of their differentiator. Naturally, the barrier to entry for people writing using AI like I do makes that field more competitive. I can imagine that is frustrating because it was already a brutal field.</p><p>My hope is that as consumers we can learn to identify the AI text generation that is helping share interesting ideas from that which does not, instead of immediately disregarding it. I look forward to continuing this experiment here in sharing ideas and takes that I think are genuinely interesting and challenging.</p>]]></content:encoded></item></channel></rss>