> ## Content Index > Fetch the complete content index at: https://bayesiantheory.com/llms.txt > Use this file to discover other available public pages before exploring further. # Life after prompting: the three architectures for using AI in 2026 - URL: https://bayesiantheory.com/life-after-prompting-the-three-architectures/ - Published: 2026-05-28T11:09:00.000Z - Updated: 2026-10-06T09:03:53.000Z - Description: Turn your prop plane chatbot into a jet fighter harness by changing your game. - Author: Gregg Bayes-Brown - Tags: cognitive, artificial intelligence, AI writing, AI, Coaching, AI training, Newsletter, #Migrated-1791211413550, #Import 2026-10-05 14:44 When I coach others on how to use AI, prompting is a foundational concept. There’s a million different acronyms people use, but the concept is simple: cover all bases, give AI everything it needs, then print outputs. While we’ve added plenty of caveats over the years, prompting as a base has held steady since 2022\. It has been the one rock in an ever changing sea. But all things come to an end, especially in AI. Over the past months, I’ve noted a massive decrease in what I’d consider classic prompting – ie. you are x, doing y, read z, give it to me with a hat on – to the point where I no longer do it at all. Instead, I just have a conversation. I’ll give you an example. A friend recently told me they were having trouble getting Chat or Claude to do some BD research for an event they are attending. Essentially, they were after a hit list which prioritises who they should be looking to speak with. Yet – despite hours sunk – neither model was delivering. I asked them to send it over. It took me 30 seconds to sort out. On my phone. In a field. Drinking cider and working on my tan while blasting out some Queens of the Stone Age at some nearby sheep. Accurate. Usable. No prompts required. Claude even threw in some bells and whistles on top like an ical mapping when all the speakers are on stage so they can grab them after. ## You can do what now? There’s been two major drivers of this change. First, the models. The models (ie. Opus/Sonnet on Anthropic, GPT on OpenAI) can now work for longer on complex tasks, enjoy a noticeable uptick in general performance, and have a much stronger ability to deliver without handholding. The second has been the introduction of the agentic harnesses. This is Claude Code and Codex (the agentic coding platforms engulfing software engineering), OpenClaw (if you don’t know, it’s why you keep seeing lobsters everywhere) and Claude Cowork - responsible for wiping nearly $1tn off the stock market when it launched. The combination of the two means AI is not only a far more competent collaborator than it was even three months ago, but is equally far superior in its ability to create all manner of reliable outputs. Critically, the jump also makes up for the severe shortcomings of agentic AI approaches in 2025. To explain, the agentic stuff we were building in 2025 is incredibly brittle - API calls stuck together with some RPA sticky-tape. These flows can be useful, but their fragility calls into question their ability to ‘replace human workers’. Now, people are capable of using agents without even realising they are doing it – with the harness doing all the heavy lifting. This turns the focus from brittle API robots into human-controlled agentic mech suits. This is why you’ve started to see a lot of people backtrack on the AI jobs apocalypse narrative. In the harness era, you still need pilots. However, we’re going from prop planes to jet fighters here. To be effective pilots, we need a new manual. This is all thinking I’m actively developing and would really welcome feedback and suggestion from readers – so please don’t take this as gospel. But instead of prompting, we now need to think about interacting with AI as the three architectures: cognition, collaboration, and context. ## Cognition As I’ve noted here before, AI is really a “reap what you sow” type situation. You want to walk into that AI arena like some mix between Mike Tyson and Einstein, ready to go twelve rounds with whatever model is standing between you and your goals. AI is really a system two game. It rewards deep thought, taking your time, testing different approaches and perspectives. This is particularly important with harnesses as the actual output bit is much more of a solved problem now (see context). What matters is the thought you put into the setup – and that includes what’s in your own cognitive architecture. I’ve outlined some of this in AI is Punk ([https://greggbayesbrown.substack.com/p/ai-is-punk](https://bayesiantheory.com/ai-is-punk/)) and in Cognitive Sparring ([https://greggbayesbrown.substack.com/p/ai-bayesics-3-cognitive-sparring](https://bayesiantheory.com/ai-bayesics-3-cognitive-sparring/)) – but what really helps here is an understanding of psychology, philosophy, and neuroscience. You want to understand your biases and fallacies, how habit works, and develop your metacognitive ability (ie. how you think about your thinking). This all informs the dynamic with which you engage with AI, which brings us on to the next architecture. ## Collaboration When we consider the human AI bridge, many people are under the assumption that it is a master-slave dynamic. You command, it obeys. This is where prompting started, but we’ve since seen it evolve so people are coached to treat AI as a junior colleague. It’s something I use myself – when coaching prompting, I say: “Would a human understand this prompt? If a human wouldn’t understand it, an AI won’t either.” That’s no longer necessarily true. Youtuber and blogger Nate Jones described the change as moving from a junior partner in 2025 to a senior one in 2026\. I concur. We’re not pulling a lever with a cookie cutter prompt. Instead, we’re taking advantage of AI’s increased intelligence, knowledge base, and ability to deliver - and aiming that straight at our blindspots. OG prompting was: “Do the thing.” 2024/25, it became: “Here’s how to do the thing.” Now, my interactions are more along the lines of: “I have a vague idea of the thing, ask me questions about it and work with me back and forth on it.” For that to work though, AI needs to know what you’re going on about. That’s our last architecture, and one you should already be well acquainted with (if not, here’s my guide on context). ## Context In the agentic harness era, how we do context has evolved. We’re no longer slapping PDFs into the chat box – we’re approaching it with far more rigour and structure. This should include an about folder (which includes information about you, your organisation, your brand, your writing style, your goals, your OKRs), a templates and examples folders (which include the frameworks and inspirations AI should be using to complete work for you) and a CLAUDE.md which defines the rules of play. I’ve prepared a whole guide on how to get this setup for Claude Cowork (available here), but some key point is setting up a folder space for your AI to work from. The other big change is skills and plugins which define workflows and how AI should approach problems you throw its way. These are easy to create with a bit of coaching and lay the foundations for your harness. You can also connect your harness to other apps and data sources using MCP connectors – normally just ‘connectors’ in the app now. This means your harness can directly access things like your Gmail, Outlook, calendar, Sharepoint and other data hubs, design tools, and a whole lot more. Something to be mindful of is context windows. A good rule of thumb is if you hit compaction during a conversation, you need to start a new conversation. Models get dumber and dumber as you fill up the context window. You go past 40-50%, you’re gonna get some really bad results. This is the only place you need an actual prompt. You can ask AI to wrap up a session and prepare a new one by giving you a prompt. Copy, paste, done. ## The gift of the gab With these three architectures in place, you don’t prompt. You just talk. No cookie-cutter prompts, no hours spent refining the perfect prompt. Just grab a mic and let rip. There’s still an art to the conversation. Start by exploring the art of the possible, use research to strengthen your approach, guide it into a plan, then do the thing. This can be slow and methodical, but it gets results. Do it enough, and it’s 30 seconds between swigs of cider in the sunshine. --- Thanks for reading. As mentioned above, I’m still developing my thinking here, and would love to hear from practitioners about how you’ve evolved your own practice in 2026\. Together, we might get to something we can share with others! [Subscribe now](#/portal/signup)