LordFall

Monetizing AI: Workflows, Agents & Business Models [Megathread]

50 posts in this topic

16 minutes ago, Joshe said:

Hell yeah, nice! Lock that shit down. Tell Claude "I'm worried someone could steal this. Let's explore security options"

WTF you can do this? Done.

The database is huge now >.<

Thanks heaps!


It is far easier to fool someone, than to convince them they have been fooled.

Share this post


Link to post
Share on other sites
On 10/11/2026 at 4:12 PM, Joseph Maynor said:

Everything is going in the direction of managing AI Workflows.  

Yeah, this is a good idea. AI Workflow setup and management for specific niches. That's the new model. Retainer fees would be necessary. Basically companies could hire a single digital problem solver for anything they need. One person with AI covering what used to take a small team. Generalists have a huge edge here. 


-

Share this post


Link to post
Share on other sites
4 minutes ago, Natasha Tori Maru said:

WTF you can do this? Done.

The database is huge now >.<

Thanks heaps!

Yep! Depending on the solution, you would likely need the Claude Desktop app instead of the claude.ai web app. 


-

Share this post


Link to post
Share on other sites
On 10/18/2026 at 7:16 PM, Natasha Tori Maru said:

I've shaved off 40% of the time spent on various tasks alone with this little project :D

 

Crazy right?
Super cool

Share this post


Link to post
Share on other sites

I finally started using Opus 5.5 as a co-worker to do my actual work (previously was just using it for random tasks and QA), and it's working great. A task that used to take 2 hours now takes 30 minutes, but the real win is the energy saved. 

I was hesitant to do this at first because I thought it would miss things, but I developed a system to check everything that could be missed. I'm now over-performing, going above and beyond on each of my client requests, and they're loving it. 

I assumed AI would make it so a slouch could do the same work as me, but that's not true because a slouch will not spend the time accounting for all the necessary particulars. They'll do the bare minimum and deliver AI slop. So, people who wield AI well are going to rise in white-collar/corporate environments. 


-

Share this post


Link to post
Share on other sites

I've seen some increases in my productivity with AI since it's launch. It definitely helps and I do use it often. It's actually great for various tasks. I'm just not seeing my experience match all the hype of people who seem to have replaced every task of theirs with AI and seemingly are doing every task through AI now. 

Share this post


Link to post
Share on other sites
4 hours ago, Lyubov said:

I'm just not seeing my experience match all the hype of people who seem to have replaced every task of theirs with AI and seemingly are doing every task through AI now. 

Yeah, I started off slow. At first I just had it doing things like going to a Jira ticket and downloading assets and verifying all the job materials and instructions. Then it sets up a job folder from my conventions so I can get started. Then last week I said fuck it, Claude, just do the whole thing. Came back perfect. Can't go back now. 

The latest Opus models know what you want without being annoying or stupid about it. A few months back you had to do a lot of set up to get that dialed in well. But it's now incredibly competent at adapting to you on its own. And it's proactive as hell: "Ok, that's done. I ran A, B, and C, since that's what we did last time. Want to do X next? If so, give me Y and say 'go'?  Haha, love it! 

Also, for digital work, nothing competes with Opus 5.5 and Fable 5.2. Fable 5.2 is too expensive though, so Opus 5.5 is the best daily driver.

Tip: use Opus 5.5 on High, not Max. 


-

Share this post


Link to post
Share on other sites

@Joshe Are you a software engineer?

I haven't tried Opus 5.5, however I find almost every AI model I have ever used has the same collection of flaws:

  • Poor architectural decision making — they will commit hard to the wrong architecture and then use their superhuman ability to produce endless hacky workarounds to make up for that poor initial decision
  • Sycophancy — when asked to evaluate something they try to make you happy rather than to properly evaluate the options
  • Highly unreadable code — even with properly configured code style guidelines, they always find ways to write extremely terse and unreadable spaghetti-like code
  • Excessive focus on edge cases — related to above, they often try to account for every possible eventually in their code, which sounds good but is actually a negative in most applications. They will sometimes structure an entire implementation around a theoretically possible edge case instead of taking a much simpler approach which would be objectively much better in practice

No matter how 'revolutionary' an agents' results are on benchmarks, I find they always have these three critical flaws, and so I have not really experienced any noticeable increase in performance since Fable.

If an AI company builds an agent that solves these, I will be far more concerned for my job. But these are hard problems to solve because their root cause is an intrinsic property of how AI models are trained and their limitations.

Share this post


Link to post
Share on other sites
10 hours ago, something_else said:

@Joshe Are you a software engineer?

No, but I'm far from a vibe coder. I have very high digital literacy and I understand systems architecture, storage, data models, CRUD, client/server, state-mgmt, tech-stacks, etc., and I've been building complex and interconnected apps with AI since it was possible, probably averaging 15-20 hours per week since the beginning. My freelance works involves a lot of frontend stuff. 

10 hours ago, something_else said:

I haven't tried Opus 5.5

Damn bro, you're missing out. Every SWE I follow agrees that it's objectively the best coding model, and it's not just benchmark hype, and it's not just a smidge better than Astra. The differences aren't marginal.

I highly recommend switching and reassessing. If you do, let me know how the switch goes for you. 

10 hours ago, something_else said:

I find almost every AI model I have ever used has the same collection of flaws:

  • Poor architectural decision making — they will commit hard to the wrong architecture and then use their superhuman ability to produce endless hacky workarounds to make up for that poor initial decision
  • Sycophancy — when asked to evaluate something they try to make you happy rather than to properly evaluate the options
  • Highly unreadable code — even with properly configured code style guidelines, they always find ways to write extremely terse and unreadable spaghetti-like code
  • Excessive focus on edge cases — related to above, they often try to account for every possible eventually in their code, which sounds good but is actually a negative in most applications. They will sometimes structure an entire implementation around a theoretically possible edge case instead of taking a much simpler approach which would be objectively much better in practice

These are mostly solvable now. They're just process problems.

- Architecture: you have to be rigorous. Something like Matt Pocock's "grill-me" skill. I recommend learning his AI-driven coding process to get an idea of good process design. After the grill session, he has the agent split the work into tickets for new agent sessions(keeps context window small and avoids the 'dumb zone'). It's an interesting process. I took his grill-me skill and turned it into a "grill-with-ui" skill that moves the grill session into the browser instead of the terminal: 

RsBelas.png

Also, you should have ADRs and agents should be mindful of them. You could have a review agent make sure no ADR was violated.

- Sycophancy: This is manageable with CLAUDE.md instructions and independent review agents. Check out Theo's process. Dude is annoying af but he knows his shit. He has Astra independently in the loop at a certain point to check Claude's work, and it can be tailored to behave however you want. Theo has it set up so that the ticket isn't complete until all the agents agree on the solution.

- Unreadable code: Solvable with agents in a loop. Simply spawn a subagent or something to check if the code matches your style guide. Also, 10 examples of your clean code is far better for AI to operationalize than code in a style guide. When Claude sees the codebase is really clean, it tries to write code in that way. 

- Excessive focus on edge cases: You could make this part of your "grill-me" process where you go over the unnecessary edge cases that AI will likely try to implement, and weed them out from the outset. 

To summarize, I think most AI coding problems can be solved with process design. If you haven't invested heavily in AI-coding process design, and if you aren't using Opus 5.5 and still using AI from 3-6 months ago, then I can see why these would be sticking points. I'm not a SWE, so I can't say with 100% certainty, but I'm pretty sure these are non issues with the right design.

Edited by Joshe

-

Share this post


Link to post
Share on other sites
14 hours ago, Joshe said:

Damn bro, you're missing out. Every SWE I follow agrees that it's objectively the best coding model, and it's not just benchmark hype, and it's not just a smidge better than Astra. The differences aren't marginal.

I highly recommend switching and reassessing. If you do, let me know how the switch goes for you. 

It isn't up to me, the company I work for chooses which platform we use. We used Claude for a while but we moved away because even though it was marginally better at coding at the time, the way Claude models wrote English was insufferable. Like it would use 2000 words to convey the same information Astra would convey in 100 words. No matter what approach we used to try and constrain it and make it write clear and coherent English, it would eventually ignore those instructions as its context window filled up.

I don't know if that's the case anymore to be fair. I will give Opus 5.5 a shot and see how it performs. Largely we have not experienced a notable increase in productivity since Fable first released. We have a comprehensive set of skills and agent files that have basically resulted in us getting almost exactly the same results from all models we have used.

14 hours ago, Joshe said:

To summarize, I think most AI coding problems can be solved with process design. If you haven't invested heavily in AI-coding process design, and if you aren't using Opus 5.5 and still using AI from 3-6 months ago, then I can see why these would be sticking points. I'm not a SWE, so I can't say with 100% certainty, but I'm pretty sure these are non issues with the right design.

I agree with all of this, but I would say these help... they don't solve the fundamental problems with coding agents. Especially the sycophancy and lack of long term maintenance and architectural thinking.

The best example I have of the issues with AI coding agents was from recently. I had to decide between using a library or building a custom implementation of a renderer for mapping document JSON into email-safe HTML.

I gave both Astra 6 and Fable 5.1 as much context on the project as they could reasonably need and essentially unlimited time to research, and both confidently recommended using the library. I then pointed out some of the cons of using the library (which it had also identified, I just reiterated them) and then immediately got back a "You're absolutely right! Actually, building the custom renderer is the way to go." And then I reiterated a few of the cons of the custom approach (which it also identified) back to it, and it flipped again. Repeat this 5 or 6 times. Both models just flip back and forth between recommending whatever approach they think you want to take.

Imagine you asked a colleague to research this for a few days and come back with a recommendation, and on the slightest bit of pushback on their recommendation they completely reversed it? Would you still trust them? No, of course not.

You can argue that the initial recommendation they had was the most objective, but I had my doubts the library was the right way to go and so I did my own research. It is objectively the wrong choice. It would have required us to extend the library with so many custom rules that we'd basically have to build the custom renderer anyway, but now the custom rendering logic would be in a separate project in a different language, and it would require us to host a separate service to process the JSON documents and return them, which is a huge overhead compared to a custom renderer built in our native language.

If the AI was given full control over this project it would have taken that approach and it would have worked, but the long term implications of it for our team would have been a fucking nightmare.

The models can not intuit that stuff like this is an unnecessary overhead and bad for the long term health of our project. They just don't have these instincts yet, no matter how much you teach them.

Quote

Every SWE I follow agrees that it's objectively the best coding model, and it's not just benchmark hype, and it's not just a smidge better than Astra. The differences aren't marginal.

SWE influencers have been saying this about every new model since coding agents became a thing. They profit from the hype, even if the benefits are largely marginal for the vast majority of use cases.

I agree with you that process + context + skills + agent files are most important to results, the actual model you are using matters a lot less once they're locked in.

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!


Register a new account

Sign in

Already have an account? Sign in here.


Sign In Now