Docuramming is a new terminology we invent. It refers to the process that to document the whole picture of what you want to build first concisely and precisely, then let AI build a complex system for you based on the knowledge we dedicatedly organized.
Different from traditional programming, Docuramming is an improved vibecoding process. The delivery of the docuramming is no longer the source code, but the docuram(the documents as a comprehensive knowledge base to prompt AI to build the system). Anyone has the docuram will be able to replicate everything the author has created and reproduce the outcome. The differences will be very minor for different runs. We can think the Docuramming is compiling natural language into programming language then the interpreter or the compiler will convert the programming language into machine instructions.
We picked a very very complex task to demonstrate the power of Docuramming. Wolverine OS is a complex financial engineering system being crafted by several extremely experienced C++ engineers and a PhD of physics for 10+ years. It works well and is now heavily used for trading in the real commodity market. The capacity of capital been traded using the strategies running on Wolverine OS is significant.
To train a qualified Wolverine OS financial indicator creator will take 2+ years the minimum. The candidates usually are either lack of strong engineering background or lack of scientific background. The candidates have to spend years to catch up with the knowledge of those two domains to be qualified. The mark for the prudence is that whether they can build solid financial trading strategies to make money from real trading. In another word, we use his real PnL to evaluate his prudence. The qualified strategy creator totally will be less than 10 on our planet for the Wolverine OS specifically.
Our boy Sanket took the challenge to build a portfolio management strategy trades multiple assets. The delivered strategy should have a good expected return and low risk. He is a CS boy graduated from Northeastern University with a degree in Computer Science, smart and hardworking but zero experience in financial trading.
This time he is not doing it from scratch, the Wolverine OS team spent years to build the knowledge base highly comprehensive and well organized. Our boy will stand on the shoulders of the giant. If Docuramming is working, he has a good chance to succeed.
He doesn't understand the financial trading, nor does he understand anything about future contracts. His task is to build:
1. tier-1 indicators for single asset
2. tier-2 portfolio management strategy using tier-1 indicators as dependencies to invest in multiple assets
The Wolverine OS has implemented a comprehensive architecture for doing experiments(backtesting) and simulating in the almost-real simulation market. Finally, the prudent strategy will be deployed to the real market environment to trade for the real war.
What is accessible to Sanket is only the level-1 price data(candle stick). So for the day 000 and 001, he spent time to warm up, the tasks for those two consecutive days is to build visualizations for the available data and some vanilla technical indicators.
On day 002, he moved on to build tier-1 indicator on a single asset. The tier-1 indicator defined in the Wolverine OS can be either a pure technical quantitative indicator outputs fields of scalers or a strategy for a single asset.
Even for day 000 and day 001, he suffered a lot for too many misleading information from the knowledge base. The rule for docuramming constrains that he can only prompt to generate the docuram(the documentations). The coder (a single Claude Code) will be in charge of implementing everything according to the docuram, or we usually call this is to compile the docuram into programming language source code.
The original authors of the Wolverine OS knowledge base don't use AI to generate strategies nor technical indicators. Their documentations have a lot of flaws and errors.Because human programmers are using the documentations as a reference to build their own strategies and indicators. It's working all the time. But for docuramming, it's not. In other words, the docuramming methodology will require the docuram to be consistent with the reality. To verify this, we can only use the docuram to generate verifiable system features. In this case, we have to purely use the library(Wolverine OS knowledge base) and the manually written docuram as the requirements and specifications to generate the inidicators.
So after the error report from day 000 and 001, the Wolverine OS team pushed a significant amount of patches to fix the errors and improve the consistency of the knowledge base. This effort was crucial for the success of the docuramming methodology.
Sanket used the updated higher quality docuram with his revised docuram to generate the source code of visualizations. One-shot success is achieved. More than that, today the tier-1 indicator finished the backtesting for the first time. He cited the PnL chart of the newly baked naive baby indicator as the journal of today.
BTW, we built dedicated features in our software TeamTurbo to exercise the Docuramming methodology. The TeamTurbo is working like a PM tool, multiple people can team together there to build. They can either use the traditional approach or the docuramming approach.
From the next day on, we will use the TeamTurbo to build a new workspace, our boy Sanket will be the first member of the seeding team for this strategy. We will welcome all friends who are interested in this build in public challenge to join us on the workspace of TeamTurbo. You can either be a part of this or to observe how docuramming or financial engineering works closely.
We strongly believe that in the near future before AGI is widely available, human+AI will be the practical way to build sophisticated large system.
Vibecoding can be improved a lot by applying Docuramming methodology. This type of improvement will bring the democracy to a lot of creative minds. A lot of people of zero programming background can build great complex system. Anyone who wants to build, please don't get hesitated, it's time to open a new era and release your creativity. We will witness and make this transformation happen together.
The reason we pick this task as the challenge is we trust the power of market and real money. If we can succeed in this domain, it's very likely this is a widely applicable approach to other domains as well.
Vibe coders now will become the mainstream army of software builders in the near future. Instead of mocking at those ‘brainless’ crowds, to build tools or to invent practical methods helping them make transition will become an industry. Serious business. #docuramming#vibecoding
To be honest, the progammers are discriminating vibe coders systematically. It is not about the video LOL. Vibe coders are furious nowadays for being judged. IMHO, they are not less creative but yet less productive when it comes to production ready tasks. This will also change very quickly.
Wow! The Pro subscription users now can use Opus 4.5! Compare with Sonnet 4.5, it is a quantum leap! Our team have transitioned to Opus 4.5 from day 1.
Claude Code users with Pro accounts are getting an upgrade: you can now use Claude Opus 4.5, our frontier coding model.
Opus 4.5 is exceptional at complex coding tasks. We recommend trying it for your toughest problems.
Claude Code users with Pro accounts are getting an upgrade: you can now use Claude Opus 4.5, our frontier coding model.
Opus 4.5 is exceptional at complex coding tasks. We recommend trying it for your toughest problems.
The Claude Code represented AI coding agent are almost all implemented in JS. And now anthropic acquires the fastest JS runtime on our planet. The lead of the AI coding agent by anthropic is more unchallengeable. I think Google should do anything possible to reinforce gemini cli. Even though they have gemini 3, the CLI is still not even close to CC.
@bunjavascript The only thing about AI truly matters to our economy is the Coding Agent. It is generating new productivity. Anthropic is taking a smart move.
@weswinder Open AI is exhausting themselves to publish new ‘products’ while the only product is high quality model in their industry. Diversity may fools the investors once but will fail the company’s vision and future.
@joexxx81 Welcome to the fancy world of AI coding. To he honest, this is an infinite deep tunnel, you will find more as you are diving deeper into. Vibecoding is low efficiency compared with other higher level ways of AI software development engineering.
@Yuchenj_UW They blow epic bubbles and are short of funding. Once the boat starts to leak, taking stupid actions to solve their urgent problems will only make them sink even quicker.
@Yuchenj_UW I agree with LeCun this time. He is always saying the ground truth however people don’t listen to him because he didn’t build breakthroughs recently. Ilya has built wonderful things, more trustworthy. OpenAI became unbelievably near-sighted. Shame.
@kylegawley Vibecoding makes you run fast. Wanna go further, it’s not reliable. We have to do our job, even though programming is not necessarily the most essential part.
@svpino If it’s augmented by human, it’s possible. The intrinsic nature about vibecoding is to orchestrate human reachable information with the strong exercise capabilities of AI. It holds untill agent can extract information deeper than us.
@theJayAlto I can’t agree more. Learning is gaining, but for those dedicatedly designed entertaining environment, you are actually taxed by given the easy rewards. We are actually burned by those apps.
@LexnL89916 @GeminiApp If you are targeting a serious career of the IT industry, you have to learn the basics intensively. To build great things and to be the top tiers employee will deviate significantly in the future.