The Rise of the Vibe Analyst
Where It All Started
In February 2025, Andrej Karpathy dropped this tweet:
“There’s a new kind of coding I call ‘vibe coding’, where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It’s possible because the LLMs are getting too good… I just see stuff, say stuff, run stuff, and copy-paste stuff, and it mostly works.”
The term went viral, and apparently was Collins Dictionary word of the year for 2025.
But here’s the thing - why should this stop at coding?
I prefer to call myself a vibe analyst, a borrowed term as I will explain!
Background: Corporate Research Assistant - an AI powered app to source SEC data
I am not much of a Python coder. I’m a chartered accountant and MBA with 30+ years experience in finance, banking, and analytics. I have intermediate skills in R for analytics and ML use, yet I have still only formally used code in a few projects (although I will often run my own little stand-alone analyses). I would classify myself as a ‘power user’ of analytics tools and an intermediate-level R-based data scientist for when it is a good fit. I am not a software programmer apart from that.
So I have done quite a few Udemy/Coursera courses, on various aspects of R, ML and deep learning (mostly R, although some of the early and excellent superdatascience courses were ‘bilingual’).
Then in 2025 I did a couple of Ed Donner’s great Udemy courses on AI, on LLM Engineering and his AI Engineer Agentic Track. Ed is a very knowledgeable, enthusiastic and effective instructor and I highly recommend his courses.
So late in 2025, I decided I should test my newly acquired knowledge (😄) using a real/test project within an area of my expertise and so began the test project:
SEC Corporate Research Assistant
This is a knowledge base and agentic AI app to extract, synthesise and report on corporate financial information from SEC filings (10-K, 10-Q, 20-F) augmented with web searches for more recent data, with some additional graphics and analytics included. The reporting ‘pack’ consists of 5 specialist/subsidiary reports which are synthesised into a comprehensive report. All data sources are intended to be fully cited for transparency. All data is then stored into a ChromaDB/Sqlite knowledge base to facilitate further user querying. It is specifically NOT an app which might provide any investment guidance, just present what information already exists in a (hopefully) meaningful way.
It all went pretty good for an initial ‘test’ project and the app/blog and github are all freely available here, at least until the OpenAI credit I loaded runs out, or you enter your own OpenAI key in the app.
But that was really just the beginning of the story…
For after just finishing the Corporate Analysis App (to a stage), it was suddenly Christmas and we had all these frozen strawberries so we decided to start a new Christmas tradition - the Christmas morning strawberry daiquiri. But even though there is always white rum in the cabinet, alas the cupboard ‘seemed’ bare - so the next AI project was born:
This took less than a day to complete (Boxing day in Australia) in a programming language and framework I did not, and still don’t, know (Next.js/Typescript/Tailwind et al)!
Both the apps Claude and I developed are free to use online and in public repos, and I have written up separate development notes about the experiences on my blog.
However, this note is not specifically about either app, more about the process I (and friends) used to build them - what I am now calling ‘vibe analysis’ and the role of the ‘vibe analyst’, both being borrowed terms as will be explained.
Also, as new tools and methodologies are coming through on pretty much a daily basis, this is where I am at the moment, but hoping to continue to keep moving forward
Beyond Code: The Vibe Analyst
So when I started ‘vibe coding’ it was by necessity - AI was a lot better coder than I will ever be, especially in Python. I did mostly look at the code (review is a bit of a stretch) if only for educational reasons. I did review the functionality while Claude, wearing his appropriate ‘hat’, reviewed the code.
But I am pretty good at finance and analytics, and so that was where I could add value, and I quickly realised that Claude and I were a pretty good team before we even got to the coding bit, if a coding bit was even needed.
Then I saw an article from the MIT Sloan Management Review from July, 2025 “Vibe Analytics: Vibe Coding’s New Cousin Unlocks Insights”. https://sloanreview.mit.edu/article/vibe-analytics-vibe-codings-new-cousin-unlocks-insights/
So the term vibe analytics has existed at least since July 2025. So I have borrowed it.
The MIT approach seems to focus on leaders engaging directly with existing data sets through conversation - getting insights faster by eliminating the translation process between business questions and technical analysis. Think executives dropping CSV files into ChatGPT and asking questions.
That is not exactly how I envisage it for my purposes, after all someone still has to do the detailed analysis. In my world I need transparency and generally some sort of audit trail. But the two approaches could happily co-exist. In any case I am happy to declare myself as a full vibe analyst - at least wherever possible.
My Approach: Building the Tools as Part of the Vibe
Where MIT Sloan describes vibe analytics as consuming data through AI conversation, I see the real power in creating the analytical infrastructure itself through the same improvisational process. And here I am talking analytical infrastructure, which is often only for a one-off but replicable, analysis. It is not intended as a operational system development, although it is sometimes useful as a prototype.
When I built the SEC Corporate Research Assistant, I didn’t just use AI to query existing data. I used Claude/s to:
- Research the best approach to extracting XBRL financial data (well with Edgartools fantastic assistance, especially after we added the related Claude skill)
- Design the knowledge base architecture (based on what I thought we needed functionally, and given some of the things I had picked up from Ed’s courses)
- Build the multi-agent orchestration system (based on the OpenAI SDK and starting example)
- Create the specialist agents for financial analysis, risk assessment, and synthesis (expanding the OpenAI example as required)
- Develop the visualisation components (still working on this)
- Deploy the whole thing!
The tools themselves emerged from the vibe. I never touched the code directly. I planned, approved, reviewed - Claude/s did all the heavy lifting.
This is fundamentally different from just “chatting with your data”. It’s about using the same improvisational AI dialogue to construct bespoke analytical systems that can provide a replicable analysis.
As an analyst, this is what I have been doing for years, except now I have some new toolsets and methodologies to much simplify that process.
A Concrete Example: Designing the Risk Agent
Let me give you a specific example of vibe analysis in action.
When designing the Risk Agent, I didn’t write prompts - I described what I wanted to see. SEC reports have been a great source of information to me over the years mainly to better know potential partners, so I have ploughed through quite a few 10-K reports.
“The risk section should capture the Item 1A risk factors from the 10-K, but also look at MD&A for management’s own risk commentary, check recent 8-Ks for material events, and synthesize it all into something an analyst would actually want to read. About 800-1200 words, with proper citations.”
Claude took that description and iterated on the prompt (I only added a very little to the prompts, Claude mainly handled them, sometimes with the Anthropic Prompt Improver) until we had something that consistently produced quality output. We went through maybe 5-6 versions (in a few agents we still have a prompt selector config), with me reviewing the actual outputs against companies I knew well (banks, mostly - my domain expertise). However this is an ongoing exercise really, at least at the moment. AI is a stochastic anaimal, we really need deterministic outcomes, so care is needed.
The same pattern repeated for every agent. I brought the domain knowledge (“what should good financial analysis look like?”), Claude brought the technical implementation. Neither of us could have done it alone.
The Practical Reality
Recently, I’ve also been using Claude Opus 4.5 to research analytical project ideas:
- Rationality check: Am I off my tree? I thought sentiment analysis was a great idea in a particular use case, Claude not so much - discussions are ongoing.
- Practicality assessment: Can it be done? (It always can - just degrees of accuracy/ levels of confidence)
- Literature review: What have others already done? This saved me from reinventing several wheels, and provided great guidance,
- Data mapping: What data exists and how do we get it?
- Methodology selection: Best analytical approaches for the problem
We haven’t even reached the coding part at this stage, if it is ever actually required.
What Went Wrong (Because Something Always Does)
The vibe analyst approach isn’t without its disasters. A few highlights from my journey:
The Great Regression of November 2025: Claude confidently “improved” the financial metrics agent and broke the balance sheet verification. Boy, Claude can move fast when deciding to go rogue! Assets no longer equaled Liabilities plus Equity, and a lot more. For a chartered accountant, this is roughly equivalent to a doctor forgetting which side the heart is on. We now have mandatory backup procedures before any changes, and stringent review/approvals steps.
Analysis Paralysis: When you get hit by a regression type disaster, you can tend to go conservative, and if you want Claude to knock out an implementation review or 20, he can do it in no time - but then you are meant to read them, apparently. So there is a happy medium between letting Claude loose and being overly risk-averse. The newer Ralph Wiggum approach (bash it until it works) may be a solution!
The Cost Explosion: Early versions of the risk agent were consuming 85,000 tokens per analysis - about $0.50 just for that one agent. Turns out Claude was helpfully including the entire 10-K risk factors section verbatim. We got it down by tens of thousands of tokens through better context management. Context management is a good thing!
The Hallucinated Citation: Once, the system confidently cited “Apple 10-K FY2024, page 47” for a revenue figure. There was no page 47. This is why we now verify every numeric claim against the source XBRL data, not the text. AI’s are stochastic in nature, so use techniques to make them as deterministic as possible!
These failures taught me that “vibe” doesn’t mean “careless”. The 4-hat approach (plan/approve/build/review) emerged directly from these disasters. I am not sure how long this approach will survive, given I am the slow cog in the loop, but it’s ok for now.
And, it’s a learning curve after all!
By The Numbers
Some metrics from the project:
- Development time: end October ’25 - Christmas ’25, but amongst other pursuits.
- Lines of code written by me: 0 (literally, maybe a few prompt edits)
- Companies analysed: 25+ in the knowledge base, with each run many times - there is nowhere to hide AAPL!
- Analysis time: it takes about 5-8 minutes to run full set of reports, but some caching is used if possible, and reports are populated gradually in the app as generated.
- Cost per analysis: ~$0.30, and we will get it much lower using different non-frontier models etc.
- Claude Max subscription months: 3 (and counting!)
- Times I nearly gave up: no, it is actually quite fun, really!
- Regressions that broke everything: a few issues, but git is always there.
The Vibe Analyst Manifesto
For finance professionals, analysts, and anyone who works with data - this is our moment. The barrier between question and insight is dissolving. Not because AI has all the answers (he often thinks he does), but because AI can help us build the exact tools we need to find them.
The person asking the smartest questions can now build the answers.
That’s the vibe analyst.
But let me be clear about what it’s not: it’s not about abdicating responsibility. I still review outputs. I still catch errors. Still make sure all the checks and balances that have always been required are still employed. The AI is a force multiplier, not a replacement for expertise.
And the one very clear ongoing constant is that data quality is essential, always (yes, I know, I spend most of my life cleaning it as well)
The vibe analyst knows their domain deeply, thinks critically about outputs, and treats AI as a collaborator - not an oracle.
And vibe analytics can be used in most areas of the finance function. I have been talking about programmed apps, but AI is now well integrated into the Microsoft Office ‘stack’ so will be increasingly available to short cut finance function processed. I will most likely write another blog soon about current toolsets which work for me and are required of me in my endeavours, ie generally assisting corporate finance units or doing corporate and project finance feasibilities and analyses - it’s changing, fast. But obviously a bit different depending on what you do.
…but who knows what Claude will be doing next week - maybe I will be demoted to the team coffee maker.
Building something with vibe analytics? I’d love to hear about it. Find me on LinkedIn or check out the SEC Corporate Research Assistant to see vibe analysis in action.