> Though it's being argued that you could still be responsible for it without looking at it?
If you have a much test coverage as SQLite, I'd be tempted to trust it, and I think that's where we are headed. Someone still needs to define these guardrails though, so IMO the developers job is just mutating into something else, but is still necessary. Long term, who knows.
Models also still can have major task adherence issues, so even a perfectly well scoped, well considered prompt that covers all situations and coverage can, in many scenarios lead to output users may not expect or notice without reading code. Seen it a lot, just lost some blood and tears to a very straight front end implementation task that I ended up applying manually cause both Sol and Fable struggle to copy a few dozen HTML/JS/CSS snippets properly into a slightly obscure codebase, despite the reference being fully featured and even using the same token scheme, etc…
We also see such issues in what labs deliver in their own software products. We are not at the stage where a human shouldn’t sign off on code that they have tested and understood. A hypothetical future, I don’t really care for, we’d notice when labs stop shipping easily apparent bugs, though even then, I’d still feel there is a while in between that and not needing to assess the output anymore.
There is, in my mind, no need for this “narrative shifting”, overtone window (don’t get me started on the idea of applying that to what should be a purely technical discussion), convince the unbelievers nonesense that some spend their time on. If models could provide code that didn’t necessitate review, there’d be no one needing to argue that case. If and when they can, we’ll know it, but not because of those espousing this currently.
Also I suspect you overprompt, with GPT-5.6 (sol, terra and luna), a lot's of what's used to be mandatory in the AGENT.md is redondant at best and frequantly actively harmful.
I do not have an AGENTS.md in any of my repos at all, stopped using them around 5.2. I was a bit ahead of the curve with that, my evals very early showcased how limiting those tend to be, same why I never used MCP, were useless context bloat.
Will maybe make an edit explaining a bit further what the task was once I get to my computer.
So this "thing" is mainly a Firefox fork and most UI is basic HTML/CSS/JS (as is the case in upstream). Development is patch baked, CSS tokens must follow a defined and CI enforced standard, etc. LLMs can be very helpful in development, I got a small CLI tool for patch, token management and basic quality gates, which I started working on a few months ago to keep the most atrocious LLM output at bay. Has lead to the revieability of output improving meaningfully over markdown monstrosities, though OpenAI models still manage to sneak hard to parse output past it. This CLI tool along with some task specific scripts also ensures reuse of proven upstream infra like Places (Good lord dear Firefox developers, is Places nice to rely on), consistent regression testing (especially in memory constraint scenarios), etc. Basically, I can and do regularly make additions with LLM assistance, I review it, I discard and restart or improve upon it (rarely accept scoped changes wholesale. This to say, I got some experience in the use of models for coding assistance and I (thanks to the amazing docs and a lot of considerations for the architecture I want) do know what I want, how I want it and how to get there. Also got private LLM evals that often uncover which labs tend to perform suspiciously well in public benchmarks vs private ones and what models still struggle with along with why, so yeah, certainly can always improve but I got, I'd argue, enough of an idea to where my critique of LLM coding limitations has legs.
Which brings us to what I was trying to implement and how I went about it: Settings works. Fully featured (including a few cross-site-tracking specific clarifications that came from a HN interaction a few days ago), tab specific previews for what changes affect regarding themeing, well tested (manual and static), integrated to leverage what FF provides where possible.
It does (or rather did) look functional/God awful though. To the point where I was uncertain that certain previews could be easily parsed by new users. I thus opened Adobe XD, did some early mockup work, tried a few core concepts, settled upon two, then (using Claude Fable 5 low) created a plain export of the existing settings code from our furnace components and patch baked edits into regular HTML/JS/CSS files. I manually verified, this export worked, the tokens were in the correct format, the code reflected what Hominis applied (including what was required for stand-alone of course) and externally called features upon interaction did provide log output linking to the pre-existing functions that meant reimplementation based upon this should be easy.
I then took that to Claude Design using Fable 5 on High. I provided the code along with linked branding files (which due to the way branding patches are handled were simpler to provide separately) and my Adobe XD mockups. A few dozen iterations later, along with some exports and re-imports due to manual changes (some animations in tabbing/"focus mode" showcases needed to be "just so" and prompting would have been inefficient to get there), I had a new user experience I was far happier with. Simpler, yet better at communicating, far more visually appealing and resolving some concerns I had, I felt pleased and will admit, Fable 5 via Claude Design provided valuable output and did, what it does best, make iterating on multiple UI concepts next to each other to settle on a final option from many, far quicker.
I then exported and took that to GPT-5.6 Luna (I have € 23,- Codex only so am a bit stingy on when to use what). But so what? I had verified, the tokens were the same. The naming of elements remained consistent to what Hominis Settings used, the backend changes were practically none-existent. I had audited the output end-to-end, made some refactors and house style specific improvements to keep everything more auditable, everything seemed suited for a quick port. What could possibly go wrong?
Anyone whith pattern recognition will likely guess what. Basic 1:1 applying? No dice. The first attempt failed as, once the context window had compacted twice, the model started leaving the very clearly paved path laid out. Stylised favicon in the showcase? Gone. Hamburger menu in the showcase, compressed. Vertical tabbing change interlinked with the canvas section? Very funny. The model started no longer following the code, it started taking screenshots and applying what it could see from that, despite the original prompt (just checked) vey clearly stating a simple code port, section per section, with any deviations to be listed in a designated file I maintain for long running tasks.
Basically, Luna did implement changes to the settings that felt tangentially right and a casual observe may not notice all the regressions and deviations, but I did. So I stopped it.
Sol and Fable didn't fare much better. Sol did stay on target longer, but it went off the rails around the privacy tab, introducing functional regressions to the way I had implemented cross-site cookie blocking, which were never requested, nor should that code even have been looked at. I reset the repo and handed it over to Fable 5 (medium). I had a third of my weekly usage left on 20x Max, reset the day after at 3AM so no harm either way.
Should be plenty. Wasn't plenty. Since a while (I think Opus 4.7, but could be wrong), Anthropic models do decently well regarding long term, high token tasks. Up to 450k, I have been able to reliably reproduce consistent implementation. The model, using a few subagents (which should have reduced the risk of context window issues further), went to work and after a few hours (and about 20% of usage less), the model proudly presented its work. I was at work and by the time I came back, I was a bit miffed to find that the model had, in its wisdom, decided to not used the well established and consistently used mar to bind in branding icons. No biggie, easy fix, albeit a bit stupid. ESPECIALLY SINCE I SAW IN THE CLAUDE CODE TRACES THAT THE MODEL HAD SURPRESSED A WARNING ON THAT VERY FRONT. Whatever. Then I saw it had not wired in the existing browser data deletion and export logic. It hadn’t modified existing logic unlike Sol, so hey, that’s nice. But it had not wired up the existing settings when they did not have any immediate feedback in the implementation reference.
Ox Alpha, it just spanned in circles, didn’t seem to like our fireforge CLI and furnace componets, but it was worth a free try. Opus 5, the model most obsessive in checking its own work, took screenshots. A lot of sscreenshots including every few hundred ms to cover animations. Nice. BUT IT CREATED ITS OWN TOKENS INSTAD OF REUSING WHAT WAS PROVIDED. Thus, styling deviated heavily.
At this point you might ask why I don’t do it manually and I will in the end anyways, but I was surprised to find such a clear case of a seemingly straightforward task flummoxing multiple LLMs. This is aided by my unique code base (the upstream FF code is also gitignored which likely flummoxes some models trained heavily to leverage git to track changes), everything needs to be patch backed and follow a specific implementation style, etc. But I had more important things to do and I wanted to see whether I couldn’t get it to work yet.
Inspired by Opus 5, I wrote a new prompt, specifically laying out a visual comparison and code diff workflow. Only these changes, only in this manner, only move on ones you have gotten visual confirmation, specific cross checks. I included a hand written markdown outlining which change affects other settings sections (even though that is obvious reading the reference code), how to approach tokens, etc. Obsessively descriptive and (I feel) unnecessarily so, but why not. Best case, it works, worst case, I’ll spend an hour doing it manually. I had other things to do not behind a keyboard, so why not one last Hail Mary.
Fable 5, ever efficient when using visuals, used the last rest of my usage, though I did see some roundabout approaches after the fact that make me doubtful it’d have cracked this. Opus 5 went off the deep end taking ui-captures across the entire code base, which lead to a very liberal application of settings tokens outside settings.
Sol did take a night and got 40% there when I asked for a pause once the in flight slice had landed. It did port the UI/UX changes in a way that on the surface looked and felt correct. It did not touch the backend in unacceptable ways. And it did cross checks. Animations also behaved correctly, though it did apply a rule on backend usage a bit to strictly, incorporating that into a preview for search by turning that into an actual web search, not a UX demo. Dumb, but not fatal.
Great success, what am I complaining?
Well, the code. It had done what Sol likes to do and turned very cleanly written, readable code into a hard to parse mess. This included touching existing test files.
And at that point I said “fuck it, I’ll do it myself”. And I did. In less than an hour, listening to Paris Palamo, Lyre Le Temps, Sting, Sade, SynthV and some Nirvana.
If I didn’t look at the code and I didn’t have strict standards for the UI, but just considered what looks in line on the surface level/feels right/“vibes” and what “works”, many of these attempts would have been accepted, as their issues are rarely apparent on the surface. That’s part of the issue in my book and why I’m firm we are far from “don’t read code”/“don’t test”/“skip qa”…
You can still write non-performant code even with tests, but you can also test performance!
What I will say is that I feel valid in saying that the developer behind gas town (a shudder runs down my spine even having to think about that) is not working in a test suite anywhere comparable to SQLite, though in fairness, you never claimed that.
Again, you are likely not wrong that with an environment like SQLite, bugs may be hard to sneak in, but then again, isn't retaining that so important that it is a strong argument for developers auditing LLM output even more strictly, lest we loose the resilience we currently have?