My current approach which I've been testing on two MVPs with what I would call 'moderate success' (but hey, actual success!)
3 tier, philosophy-spec-design. Increasing detail. Design files include db model explanations and pseudocode/function headers - that level of detail.
For each thing I need to change, I have a, prompt ready to go to ask the agent to follow about 5 steps and it outputs a 'reviewfile' with details of what it things about the thing I posited. I review its output. I have another prompt ready to then get an agent to generate a taskfile + update the design documentation. The taskfile explains in great detail what has changed and what needs to be implemented. I review the taskfile and got diffs of the design doc changes. Finally an agent implements the taskfile. I review all changed code and commit.
It gets there, but still definitely misses some stuff. It's very adequate for a MVP I'm finding.
Edit: this seems to only work with Opus. Sonnet can't do it (maybe I'm just lucky and Opus is seriously compensating for an awful approach and I'm just lucky?)
I had to do the same. I expect everything will go token pricing, and at that point a LOT of small/mid businesses will drastically change how they use code.
I've swapped to the 20x Claude plan for a month or two to knock out two ideas I need to get it MVP - expecting Claude to go token priced soon.
I have spent a SOLID 3 full days 8h/day (plus long running tasks overnight) thrashing out a random idea for a Web application using purely Opus (mostly Max, sometimes ultracode version). I'm not a project manager, but I genuinely tried a full 3-tier spec out - design->specs->build details.
While it was significantly better than previous attempts, it still misses very basic things - sporadically. Eg. A clear design requirement was essentially adding clients, explained clearly and comprehensively. The ability to add clients was entirely missed in the build and iteration (there were multiple 'please check its all done' separate agent runs/checks).
I can imagine in a fully autonomous deployment, in even moderate complexity, even to this day would still occasionally mess up - badly enough to cause non-trivial business issues.
I haven't managed to really figure out what's the best way, but my latest thinking is really having boil down tasks to almost unit operations "add UI button, wire to Api call. End".
Unless the token estimates I get from using Claude are wayyy out, I burn through 5m+ tokens/day, and I'm not doing a lot of time. 500k tokens in a 24h period for $5k of hardware seems quite poor?
Hey I mean we've been using a fairly basic mssql box for years in a organisation of about 300 and it humms along nicely - granted we have 2 'materialised' aggregate tables/views (done in Python) but the TCO is tiny, and there is very little learning curve.
Everyone knows a philosophy comes before a spec. Claude has to write your applications philosophy first, then you write your spec. But a philosophy is crap without a values statement, so Claude has to actually write that first.
I'm really annoyed at github for only allowing export of a single month of usage.
There were a lot of people (here included) that were absolutely abusing github - getting Opus to generate its own subagents for days on end with a single premium request. If THAT'S $3k/mo, I'm honestly not worried.
I was just asking sonnet, 5.4, opus, in single agent session to fix a problem for 20 minutes...if THAT'S $3k/mo, then AI is truly cooked.
I've so far been anti-chinese memory in my recommendations, not because it's Chinese (I don't really trust any big organisation/govt other than the EU?) but because they've been very new, and it's not worth PC stability for saving $50.
However with corsair giving it their blessing, and their technology having matured a bit (a lot?), and more reviews showing good stability (longevity I suppose, is TBD) they're definitely worth recommending these days.
It absolutely is in Australia, and I suspect Europe (from when I went 5 or so years ago). Even mostly possible in Houston when I was there 15 years ago.
The US is even more cooked than I thought if you can get a bunch of fresh veggies from a grocery store.
I think art/music is in a weird dichotomy of sorts, for me at least. It's both the most and least impacted by AI. It's trivial to go "hey suno give me music" but I refuse outright to listen to AI music as knowing it's statistical noise (of sorts) I don't really get value/enjoyment out of it.
I'm confused how the world itself works - or would work - at scale. Taking that minecraft clip on the link, if I jump down a cave first, and the model decides the cave goes off to the left. A few hours later I come back, and this time the model decides the cave bends to the right?
The only way you can get away with creating an application without touching sql is of you offload the logic to your backend language, and then I don't think you'd be efficient enough to scale.
Also can someone actually understand the logic of joins, indexes, pks, etc enough to create an efficientand scalable db, and not simply have learned sql by proximity?
Copilot has said they'll be giving out previous months "if you did token pricing, it would cost $x" so a lot of us will have real numbers to actually anticipate our spend.
Personally I'm anticipating agentic coding will be out of my price bracket (a single agent run costing US$20+ is far beyond what I can justify, especially with how often it fails). I'm planning on going back to optimised prompts on one-pass edits.
>there's little correlation between token spend and the quality
My sentiment exactly! I have a very similar scaffold to each of my prompts, and feel I provide similar context files, however sometimes I get a truly inspired, complex, and functionally complete response...and sometimes I'd have been better off running lorem ipsum through a python interpreter.
I can't find any rhyme or reason to success. I'm not sure if prompting is significantly more nuanced than I realise, or it's the statistical magic that's having a laugh at me.
>open source models on inference-optimized hardware.
Is this actually a thing? Or are you talking about some hypothetical "opus 4.7 ASIC"?
3 tier, philosophy-spec-design. Increasing detail. Design files include db model explanations and pseudocode/function headers - that level of detail.
For each thing I need to change, I have a, prompt ready to go to ask the agent to follow about 5 steps and it outputs a 'reviewfile' with details of what it things about the thing I posited. I review its output. I have another prompt ready to then get an agent to generate a taskfile + update the design documentation. The taskfile explains in great detail what has changed and what needs to be implemented. I review the taskfile and got diffs of the design doc changes. Finally an agent implements the taskfile. I review all changed code and commit.
It gets there, but still definitely misses some stuff. It's very adequate for a MVP I'm finding.
Edit: this seems to only work with Opus. Sonnet can't do it (maybe I'm just lucky and Opus is seriously compensating for an awful approach and I'm just lucky?)