Author states, “Single developer projects can build to the caliber and consistency of large development teams.”
When I, as a single person, can produce a project in one month that would have taken a team of four people three months to produce, why would I care about token cost? I’m now spending $500/month instead of $40,000 month to get the same thing 3x faster. $500 for a project instead of $120,000. (Assumes my cost, $40k is the other three people)
Have you laid those people off? You still have people on the payroll. If each one of them uses a $500 subscription, you are now spending $42k/month, not $40k.
Reading this really makes me wish that I had a slightly better workflow. I'm really soley dependent of Fable to the point that I don't use other models, and I've already sort of hit a point where I'm running into usage limits every week. I am really living on borrowed time — when Anthropic finally collapses their 50% usage increase at the end of August, I'll definitely be forced to switch my workflow. When that happens, I have a hard time imagining that I'll be sticking with a single model on a single provider.
> Reading this really makes me wish that I had a slightly better workflow. I'm really soley dependent of Fable
Good thing then that you're fully in control of your own actions.
A gift granted to us by being a fully grown adult that is also likely registered to vote.
Meta: There is of course a way to put this less snarky, but that doesn't slap people in the face as hard as they need to be slapped in the face to maybe one day start remembering that they have agency.
Once it has exactly your coding conventions and access to other code to copy bespoke patterns, a strong idea for what to do, then you can let it do the work.
I've been using GPT-5.6 Luna for everything at this point. It is more than enough to do all the work I need to do. You have to lead it by the nose, but if you know where you're going it works really, really well.
I’ve just been using it, with a chat interface. Basically, as a “consultant.”
For me, and my projects, it’s been great. It’s made an enormous difference.
I guess my workflow may seem “quaint,” to many folks, here, but the end results speak for themselves.
I suspect that one vocation that could get heavily impacted by AI, is the consulting business. That’s where many experienced people go, as they reach their career peak.
In my last project (just about to ship), ChatGPT replaced a whole bunch of services that would usually be supplied by external advisors.
But these are also services that I would normally not be able to afford, otherwise, and would just have to “make do” with. This release will have a level of polish that I have would never been able to achieve, unassisted by AI (I had originally used “on my own,” there, but the reality is, it actually was “on my own”).
It seems there are two distinct groups, one which is doing relatively well with Luna, Deepseek Flash and similar models while other seem to be satisfied only with the state of the art.
Depends on how in-the-loop you want to be. I personally delegate all the writing of the code to agents, but I maintain a clear mental model of the architecture, which I come up with by iterating and prototyping with agents. I can use Luna for all of this, although I switch depending on the task. It's nice to be able to throw a 1,000 word rough spec at Fable and get a personal tool that works perfectly though. I expect that as the models get better, I'll continue to be in the loop, but operate at increasingly higher levels of abstraction.
Having worked with both, it seems like a control thing to me. Either you're cool with Fable spitting out tons of code you'll never read, or you're cool with Luna doing targeted work while you manage the main work thread.
(I'm in the second boat so long as I'm responsible for the code I PR)
I'm ok with not-Fable until I need anything design-related: a nice HTML page, LaTeX typesetting, UI design. Sol is especially incapable of doing anything sensible.
Is there any alternative model with design sensibilities?
Did you add some skill for this? I've been using a frontend design skill with Sol and it's reasonable (for my purposes). I don't remember where I got that skill.md though, it probably was from OAI's own blog about this a few months ago.
I'm happily in both depending on what it is. Even Sol / Fable cannot do some truly novel stuff and if you rely on it too heavily you get detached from the underlying systems to the point that it's both uncomfortable and detrimental.
Sol is basically at the same level, but with better usage. You're not dependent on Fable. You could switch the next minute and hardly experience quality differences.
My experience of Sol is that it's very instrumentally competent but is a complete terrorist that will cheerfully do unhinged things that "technically" solve a problem. For example, I asked it to verify something in our integration tests and it responded by grabbing kafka details out of config and trying to use them. When I pointed out that this was not it, it did the same thing with a database client instead. Conversely, I'm pretty sure Fable or even Opus would have done something consistent with what the rest of the codebase does (consume an API or point out that there's no API to access in this case).
I don't have Fable at work but I'd probably use it for actual code if I did because not having to spend time handholding the model on this stuff and getting useful code first try is very useful
Fable is pretty good at routing to Opus and Sonnet though, especially if you ask. also, sometimes just because it decides that's what it feels like doing at that point.
What I've learned in last week is that a harness is basically a while loop.
In each iteration you make an LLM call, perform some work (e.g. tool call), augment the prompt (append or compact etc.)- not necessarily in that other- and continue.
Until an end condition is satisfied. Then you break out.
I've been very happy with Luna but my approach is "many bite sized edits" for which models basically hit saturation a year ago.
(I also tried the "let a massive model make massive changes" approach and am still psychologically recovering from the experience. The codebase may never recover!)
Also, Luna and DSV4 Flash seem to be on par now except Luna is faster and cheaper?
FWIW I run a clinical analysis backend and directly compared Luna with DSv4 Flash 0731 - DS was a bit ahead, and cheaper even considering it used 40% more tokens.
The exposure to Deepseek made me question the valuation house of cards built on SOTA providers. There are more companies producing competitive and useful models than there are companies producing jet engines for airliners, and not for the lack of trying. China has been trying to make these engines for decades and so far failed (their flagship C919 airliner is using CFM, American/French), but it has produced at least three competitive model companies within 3 years even though they're handicapped by their hardware.
It's simply not that hard, and diminishing returns will, in fact, diminish.
LoRA adaptive learning using open-weight models and your own reasoning traces is the thing. The big labs have a mammoth job ahead of them if they want to compete with running your own model - they will basically have to give every single user their own persistent virtual machine. When it's all said and done, I think their only really moat will be as inference/hardware providers. Stripe buying OpenRouter was a very smart bet.
You train an addend to the weights rather than the weights themselves. You limit the dimensions[1] available in this addend to a small fraction of the dimensions of the weights, making training cheaper.
When you have a few example chats you want a model to emulate - say you made it from your proprietary data, you can train any open model on that data in this cheap way. You don't lose any quality versus not using lora since the models overall knowledge won't shift that much due to your data anyways, so it's a waste to make high dimensional updates.
However, only in some cases is it worth it and equal in quality to just making a good retrieval system and exposing it to claude code or whatever. If a retrieval system over the same data is very difficult, or if the data simply must be proprietary, then you should go for it.
I use Claude Code, Codex and opencode pretty much interchangeably. I am currently using Claude more this month because (stupidly) I paid for Max ($100) since I have a large client project.
I barely spend time or thought in making prompts. I have a simple text file with a few combinations. They refer all the common files (README, AGENTS, DEVELOP, etc.)
All reference software is cloned locally and the docs mention that. The prompt templates then boil down to research mode (write Epic) or worker mode (write software) or review mode (leave review notes in Epic). That's it.
Many of my harness experiments are about text manipulation, text search, graph on text. Because that is what LLMs are - text processing systems. Cut parts of prompts, cut parts of response, cut parts of user's intent. Join, break into epics/tasks, run with LLMs, repeat.
This guy knows harnesses!! I'm curious to know what your thoughts are on dsh(deepseek harness), where everything is a plugin. I read an article where the author described it as a motherboar in the Mobo combo purchases we use to do to save money, with the processor(deepseek), being replaceable by other llms but shipped with it by default. Did you evaluate it as a possible replacement? If not, why not?? In my nondev brain, it looks revolutionary, the equivalent of the x86, to the SoTAs attempt at keeping us in a walled garden similar to apple consumers.
am I the only one here who is perfectly fine with deepseek and qwen? generally I do all the planning in my head and a few sentences describing what I want is enough for deepseek to save me hours of work. Hell I get away with flash for most of it and use pro for some reviews.
> Single developer projects can build to the caliber and consistency of large development teams.
This has always been true. Good developers, like truly good devs, could run rings around a team of mediocre devs. It’s a multiplier, a team of 10 1x devs will get dominated by a single 10x dev no matter how much AI they use. Nothing has changed here, if anything it benefits the good developers.
> At the moment my rig is supported by two subscriptions (Cursor, Claude) that I can augment with Pi as needed.
This conflicts subscriptions with an actual harness, doesn’t bode well for the rest of the article…
> Recently I learned about prewalk, Can Bölük's technique that uses frontier for the planning phase and first task, then hands off once the pattern is set.
This pattern has been known for years and is not attributable to a single person.
> Exploration leads to a plan formalized into an explicit DAG (directed acyclic graph) task list. Then a worker takes over, focusing on implementing the DAG one node at a time. Once complete, I bring in the critic to simplify and question what was implemented. Often this phase will push back enough that the worker phase is revisited. But once satisfied, the critic gives way to a promoter, which is my reminder that a job is not complete until you've properly communicated it to others.
I’ve tried all these complicated workflows. In the end the best way to use LLMs is to give it some instructions, take a look at the code, and then ask it for changes. At the end, ask it (in a fresh session) to review the changes for bugs or incorrect assumptions and architecture. Rinse and repeat. Anything more complex is over engineering.
None of the rest of the article seems particularly interesting. Just more busywork.
Is the article slightly naive? Sure. It describes a relatively uncomplicated multi-agent system (which is good). But frontier model handing off a DAG rather than a raw markdown to a cheap model seems like a decent idea. Ultimately, this is about cost saving, not peak performance.
For general workflows, I agree. Doing some DAG system is just a waste of time if you don't have a grounded verifier for every node. If you are going to human verify, why make a dag of subagents and waste time? Keep the dag in your head, the way we all did before LLMs. The dag in your head is also much better.
For specific workflows however, even this is under engineering. If your specific task or family of tasks can truly be decomposed into multiple verifiable subtasks then you should spend the effort to build that system. It doesn't matter if it's technically worse than the other method because it is basically automatic and incredibly cheap. For complex systems this is a lot of upfront work, but if it is done, then the return is completely outsized.
I painted it black and white, but you can of course compose everything.
When I, as a single person, can produce a project in one month that would have taken a team of four people three months to produce, why would I care about token cost? I’m now spending $500/month instead of $40,000 month to get the same thing 3x faster. $500 for a project instead of $120,000. (Assumes my cost, $40k is the other three people)
It’s a no-brainer —- use frontier all the time.
Good thing then that you're fully in control of your own actions.
A gift granted to us by being a fully grown adult that is also likely registered to vote.
Meta: There is of course a way to put this less snarky, but that doesn't slap people in the face as hard as they need to be slapped in the face to maybe one day start remembering that they have agency.
And Opus5 aggressive audits.
Once it has exactly your coding conventions and access to other code to copy bespoke patterns, a strong idea for what to do, then you can let it do the work.
You do the wiring, it fills it in.
Coding was never the work.
For me, and my projects, it’s been great. It’s made an enormous difference.
I guess my workflow may seem “quaint,” to many folks, here, but the end results speak for themselves.
I suspect that one vocation that could get heavily impacted by AI, is the consulting business. That’s where many experienced people go, as they reach their career peak.
In my last project (just about to ship), ChatGPT replaced a whole bunch of services that would usually be supplied by external advisors.
But these are also services that I would normally not be able to afford, otherwise, and would just have to “make do” with. This release will have a level of polish that I have would never been able to achieve, unassisted by AI (I had originally used “on my own,” there, but the reality is, it actually was “on my own”).
(I'm in the second boat so long as I'm responsible for the code I PR)
Is there any alternative model with design sensibilities?
I don't have Fable at work but I'd probably use it for actual code if I did because not having to spend time handholding the model on this stuff and getting useful code first try is very useful
In each iteration you make an LLM call, perform some work (e.g. tool call), augment the prompt (append or compact etc.)- not necessarily in that other- and continue.
Until an end condition is satisfied. Then you break out.
(I also tried the "let a massive model make massive changes" approach and am still psychologically recovering from the experience. The codebase may never recover!)
Also, Luna and DSV4 Flash seem to be on par now except Luna is faster and cheaper?
The exposure to Deepseek made me question the valuation house of cards built on SOTA providers. There are more companies producing competitive and useful models than there are companies producing jet engines for airliners, and not for the lack of trying. China has been trying to make these engines for decades and so far failed (their flagship C919 airliner is using CFM, American/French), but it has produced at least three competitive model companies within 3 years even though they're handicapped by their hardware.
It's simply not that hard, and diminishing returns will, in fact, diminish.
When you have a few example chats you want a model to emulate - say you made it from your proprietary data, you can train any open model on that data in this cheap way. You don't lose any quality versus not using lora since the models overall knowledge won't shift that much due to your data anyways, so it's a waste to make high dimensional updates.
However, only in some cases is it worth it and equal in quality to just making a good retrieval system and exposing it to claude code or whatever. If a retrieval system over the same data is very difficult, or if the data simply must be proprietary, then you should go for it.
[1] technically "rank", but I'm simplifying
I generally use larger models to plan. All my generated Epics have similar structure. All my repos have similar structure (https://github.com/brainless/akar and https://github.com/brainless/daftprompt are recent examples).
I barely spend time or thought in making prompts. I have a simple text file with a few combinations. They refer all the common files (README, AGENTS, DEVELOP, etc.)
All reference software is cloned locally and the docs mention that. The prompt templates then boil down to research mode (write Epic) or worker mode (write software) or review mode (leave review notes in Epic). That's it.
Many of my harness experiments are about text manipulation, text search, graph on text. Because that is what LLMs are - text processing systems. Cut parts of prompts, cut parts of response, cut parts of user's intent. Join, break into epics/tasks, run with LLMs, repeat.
the pattern of layering of deterministic, probabilistic, deterministic, probabilistic. it's a strange pattern but it seems somehow natural.
> Single developer projects can build to the caliber and consistency of large development teams.
This has always been true. Good developers, like truly good devs, could run rings around a team of mediocre devs. It’s a multiplier, a team of 10 1x devs will get dominated by a single 10x dev no matter how much AI they use. Nothing has changed here, if anything it benefits the good developers.
> At the moment my rig is supported by two subscriptions (Cursor, Claude) that I can augment with Pi as needed.
This conflicts subscriptions with an actual harness, doesn’t bode well for the rest of the article…
> Recently I learned about prewalk, Can Bölük's technique that uses frontier for the planning phase and first task, then hands off once the pattern is set.
This pattern has been known for years and is not attributable to a single person.
> Exploration leads to a plan formalized into an explicit DAG (directed acyclic graph) task list. Then a worker takes over, focusing on implementing the DAG one node at a time. Once complete, I bring in the critic to simplify and question what was implemented. Often this phase will push back enough that the worker phase is revisited. But once satisfied, the critic gives way to a promoter, which is my reminder that a job is not complete until you've properly communicated it to others.
I’ve tried all these complicated workflows. In the end the best way to use LLMs is to give it some instructions, take a look at the code, and then ask it for changes. At the end, ask it (in a fresh session) to review the changes for bugs or incorrect assumptions and architecture. Rinse and repeat. Anything more complex is over engineering.
None of the rest of the article seems particularly interesting. Just more busywork.
> HUMANIST SOFTWARE DEVELOPER
Uh huh?
It's not
For general workflows, I agree. Doing some DAG system is just a waste of time if you don't have a grounded verifier for every node. If you are going to human verify, why make a dag of subagents and waste time? Keep the dag in your head, the way we all did before LLMs. The dag in your head is also much better.
For specific workflows however, even this is under engineering. If your specific task or family of tasks can truly be decomposed into multiple verifiable subtasks then you should spend the effort to build that system. It doesn't matter if it's technically worse than the other method because it is basically automatic and incredibly cheap. For complex systems this is a lot of upfront work, but if it is done, then the return is completely outsized.
I painted it black and white, but you can of course compose everything.