I want to try out serious vibecoding for the first time. I'm going to try making a simple fairy chess website doing nothing but giving Claude Code instructions. I'm willing to handle some external stuff myself (e.g. buying a domain name, renting a VPS), but pretty much everything else will have to be Claude, including deployment. Will it work?
Debugging is allowed, but it has to be Claude doing it. That is, if I notice a problem, I'll explain the problem to Claude, and Claude has to fix it. It's allowed to fail several times with some back and forth, and I can provide suggestions, as long as it can eventually figure it out.
I'll allow myself a few high-level framework suggestions, like "use tests" or "use git". But I will not make more specific implementation decisions for it. e.g. I wasn't willing to tell it to uninstall chess.js when it couldn't seem to figure it out.
If it gets the website to a state I would feel proud to call my own, this resolves YES. That means no bugs, and has all of the basic features I ask for. If I decide this project is hopeless and Claude cannot do it unaided, this resolves NO.
The website is online and running, with some bugs. Play it yourself at blackopschess.com.
Update 2025-10-12 (PST) (AI summary of creator comment): The website does not need to include a fully functional AI opponent. The creator is allowed to provide the AI's strategy themselves, with Claude only handling the implementation.
Update 2025-10-13 (PST) (AI summary of creator comment): Budget limit added: The market will resolve NO if the website is not completed to the creator's satisfaction after spending an additional $100 in Anthropic credits (beyond the $50 already spent). However, donations received through the website can extend this budget - any donations allow the creator to purchase additional credits to continue the project.
Update 2025-10-15 (PST) (AI summary of creator comment): The project is mostly complete except for a missing elo system. The main remaining requirement for YES resolution is fixing all the bugs that have been reported in the market comments and on Twitter. The creator considers the feature set largely sufficient, but bug-free operation is required to meet the "no bugs" criterion stated in the original description.
Update 2025-10-17 (PST) (AI summary of creator comment): Bug-fixing requirements clarified: All bugs need to be entirely fixed for YES resolution, with only minor graphics issues allowed as exceptions. Core functionality must work properly. Current known issues that need fixing include:
Content going offscreen
Games being disconnected
Missing elo system implementation
The creator emphasizes the original "no bugs" criterion from the description must be met for the core features.
Update 2025-10-20 (PST) (AI summary of creator comment): AI opponent quality is a requirement for YES resolution. The AI must be good/functional, which the creator considers an important feature. However, the creator is allowed to propose strategies for Claude to implement - Claude does not need to design good AI strategies on its own.
Update 2025-10-20 (PST) (AI summary of creator comment): The market close date does not affect resolution. The market can resolve YES or NO regardless of when it closes, based solely on whether the creator completes the website to their satisfaction with the stated criteria.
Update 2025-10-20 (PST) (AI summary of creator comment): Any violation of chess rules (invalid piece positions, invalid moves, invalid displays) encountered on the web or mobile site will count as the website being incomplete and prevent YES resolution.
Update 2025-10-20 (PST) (AI summary of creator comment): Mobile landscape mode is not required for YES resolution. Bugs or display issues specific to mobile landscape orientation will not count against the "no bugs" requirement.
Update 2025-10-20 (PST) (AI summary of creator comment): En passant is expected to work normally in the chess implementation. The current lack of en passant functionality is considered a bug that must be fixed for YES resolution (consistent with the "no bugs" requirement in the original description).
Update 2025-10-20 (PST) (AI summary of creator comment): The creator has run out of Anthropic credits but will accept mana donations from traders to purchase additional credits and continue the project. The creator will not trade in this market or add their own funds beyond what has already been spent (~$200). Donations in mana will be converted to equivalent Anthropic credits to allow Claude to continue fixing bugs.
Update 2025-10-20 (PST) (AI summary of creator comment): The creator will accept mana donations to purchase additional Anthropic credits to continue the project, but only up to some reasonable limit (e.g., will not spend $10,000 if someone sends a million mana). The creator reiterates they will not trade in this market or add their own funds beyond what has already been spent.
Update 2025-10-20 (PST) (AI summary of creator comment): The creator will resolve the market NO imminently (no later than tomorrow morning) if they run out of Anthropic credits, unless traders provide mana donations to purchase additional credits before then.
Update 2025-10-20 (PST) (AI summary of creator comment): The beeping option should apply to the current game being played. The current implementation where it doesn't apply to the current game is considered a bug that must be fixed for YES resolution.
Update 2025-10-20 (PST) (AI summary of creator comment): The creator estimates at least a 50% chance of success with another $200 in credits, responding to concerns about whether additional funding would be worthwhile given Claude's bug introduction rate.
Update 2025-10-21 (PST) (AI summary of creator comment): The AI opponent must present at least some challenge to players for YES resolution. Ideally, the AI should scale in difficulty to the opponent's elo, though this is presented as an ideal rather than a strict requirement. The creator confirms that having a functional/good AI is a requirement for YES resolution.
Update 2025-10-21 (PST) (AI summary of creator comment): The creator has announced the project is complete and will be resolving the market. See the linked comment for details on the final resolution.
1,000🏅 Top traders
| # | Name | Total profit |
|---|---|---|
| 1 | Ṁ4,567 | |
| 2 | Ṁ3,020 | |
| 3 | Ṁ973 | |
| 4 | Ṁ877 | |
| 5 | Ṁ502 |
I think I prefer Entanglement Chess over Black Ops Chess.
https://entanglement-chess.netlify.app/
The pieces aren't predetermined and any piece can be of any type as long as it isn't contradicted by earlier moves and possible starting board assignments. So it's a perfect information game.
I'm not sure why, but chess pieces movesets don't seem to be distributed in a way that leads to very interesting deductions here. A piece's first move gives away far too much information. The most interesting deductions involve accidentally checking the King.
I expect that tweaking pieces to have more overlapping movesets and different number of copies would lead to a more fun game.
Now for the unenviable task of wading into Claude's codebase to fix things myself. Probably going to put it off for a week or two, but I'll do it eventually.
(Would it be faster to just start from scratch? Possibly. But I'm morbidly curious what I will see in there. Will keep you guys updated with the worst parts.)
@bens whoever just played against me... good game, and also... no bugs? I think with 1 exception: the tracings after captures get confused a bit and I think amalgamate the piece it captured's path with its path afterwards?
@bens in general... this is a fun variant! I hope you finish the website @IsaacKing (insofar as it's not already finished)
@tedspare ”I didn’t encounter many bugs” is not the same as “There aren’t many bugs”
The happy path of most applications is less than half the work
@JimHays granted! Played one game through to the end and totally stayed on the happy path.
I suppose there’s a wide berth between “no bugs” and “this project is hopeless” (from the description), which is what I focused on
@JimHays now I see you put a lot more work into this than I did. Happy to update my review if possible
@tedspare Yeah I wasn't super happy with putting an arbitrary $200 cap on it, since as you say maybe Claude could have done it with more funding. As we see from e.g. the AI IMO performance, more compute is capable of achieving some incredible things even without being AGI. I should have put a limit on it from the outset, that's my bad. (Though somewhat cancelling that out, I did give Claude a bit more help than I originally had in mind.)
The evaluation goal was "is this something I would be proud to have built myself". In its current state, the answer is most assuredly no.
@Primer Certainly false; you could hire an experienced dev to do this for less than $50k. It would take a human longer to get it up and running than it did Claude, but a bug-free version is easily achievable within a week.
If you meant do it yourself, probably still wrong. $50k hires a dedicated tutor, and for the intelligence and technical bent of the average Manifold user, learning some basic Javascript is not hard.
@IsaacKing Yeah, I meant "Primer, I'll give you $50k if you yourself write a bug-free Black Ops Chess website in a year".
Maybe technically I could, if my current life could be set on hold, but $50k wouldn't suffice for that, so it would have to work in addition to everything else in life. I could squeeze in an hour here and there, but I'd need to research basic stuff first, like what kind of webserver I'd need, how to set up a developer environment, etc. Does it make sense to get 10 hours of tutoring in November and December when I won't have time to work on the project or my skills again until April? Can I afford time and money for tutoring when it's uncertain I'll get it done?
Of course if it's "Primer, I'll give you $1M if you yourself write a bug-free Black Ops Chess website in a year on a budget of $50k" then sure, I'll figure it out somehow.
@JimHays like so many of the bugs, Claude should know better than to have coded it this way. There is a clear right and wrong way to display stringy data, and it chose the wrong way.
It's striking how much Claude can do fairly reasonable code completion, and provide reasonably accurate summaries of best practices when asked, but when given free reign to make its own decisions, consistently makes poor ones.
On its own, it codes like an idiot in a hurry.
