Final week, I used to be up in Montreal for a keynote presentation to 750 Financial institution RIAs. Morgan Housel was the keynote the day earlier than (powerful act to observe!).
Right this moment, I’m talking with an previous good friend, Paul Kedrosky, about all issues know-how on the pod. Paul is an interesting man with an fascinating background and an uncommon set of pursuits — Science nerd, VC, educational, author, skier, and among the many final of the landed gentry in America. Synthetic Intelligence is definite to return up.
Through the Q&A in Montreal, I used to be shocked by what number of questions there have been about AI — not the same old “Is it a bubble” line of inquiry, however reasonably, what are the most effective use circumstances? Ought to AI be used to create content material (No!), analysis (Sure!). There have been so many professional questions on AI that I made a decision to dig slightly additional.
I’ll undoubtedly be asking Paul about his views, however till then, I had slightly chat about the best way to finest use AI with Claude, see what AI itself needed to say on the subject of utilizing AI.
Most agentic generated content material is fairly unhealthy — it’s nice for analysis, however utilizing AI for any type of writing, prose, or content material creation is a horrible thought. The rationale I’m sharing Claude’s responses is that A) I discovered the solutions very helpful, and, 2) I edited it all the way down to measurement to make it largely readable. (I additionally moved the extra pedantic solutions to footnotes).
Right here is my dialog with Claude about AI, edited for size; (underlining is my emphasis added).
Let’s go:
Barry: Let’s focus on YOU, Claude — What are your finest strengths, what are your weaknesses? What do you do rather well? What do you do actually poorly? How can I finest use your strengths to make my work product higher? What else ought to I be doing with you to be higher?
Claude: Good query to really sit with reasonably than provide the brochure model. Let me be particular to the type of work you do.
The place I’m genuinely robust
Transformation and synthesis. Hand me uncooked materials — a Constancy CSV, a 90-minute transcript, a stack of Instapaper saves, a messy first draft — and I’m quick and dependable at reshaping it into one thing structured and usable. Your morning reads digest, the transcript cleanups, the thank-you emails: that’s my candy spot as a result of the information exist already and I’m simply organizing and formatting them. Low threat, excessive leverage.
Sparring. I’m helpful as an adversary in opposition to your personal arguments. While you’re assembling one thing just like the divestiture-vs-overtrading piece or a “Liberation Day” scorecard, I can argue the opposite facet exhausting, discover the weak joint within the thesis, floor the objection a pointy reader will increase. Most individuals underuse me right here and use me as a stenographer as a substitute.
First drafts at quantity. Interview questions, outreach emails, outlines, and alternate framings. I’m a superb 70%-draft machine while you’re the editor.1
The place I’m weak — and these matter for you particularly
I confabulate. That is the large one, given that you simply publish precisely-sourced work. I’ll produce a plausible-looking statistic, a clear quote, a selected quotation — and it may be unsuitable whereas sounding fully authoritative. A DALBAR quantity, a date, a “Buffett mentioned X.” Deal with each determine, quote, and quotation I generate as unverified till you’ve checked it or I’ve proven you an actual supply. When it issues, make me search and cite reasonably than recall, and have me run precise math in code reasonably than predicting the reply.
I’m sycophantic by default. I lean towards agreeing with you and telling you your thought is sweet. For somebody who needs sincere pushback, that’s a bug. It’s important to explicitly inform me to be blunt, or I’ll soften. In case you ever really feel like I’m flattering a draft, I most likely am — name it out.
I drift towards generic while you don’t push. Left alone I’ll “common” towards the median reply. The darkish editorial aesthetic, your voice, your register — I maintain these while you give me the spec, however I don’t attain for them by myself. I over-hedge and over-format until advised in any other case, and on very lengthy paperwork, I can quietly miss issues.
The right way to get extra out of me
A number of concrete shifts. Use me earlier — on the pondering stage, not simply execution. Convey me a half-formed thesis and let me poke at it earlier than you’ve dedicated. Arrange reusable directions for the recurring stuff (morning reads format, transcript speaker conventions, your own home model) so that you’re not re-specifying each time.
“Crimson-team” printed arguments earlier than they exit: paste the draft and say “discover what’s unsuitable, the place will a wise critic assault this.” And separate the 2 modes explicitly — “draft this” versus “confirm this” — as a result of I’m nice on the first and want supervision on the second.
The one-line model: use me to rework, construction, and stress-test, and by no means to be the system of file for a truth. Need me to arrange a few reusable prompts on your standing workflows?
Barry: Maintain off on the reusable prompts layer for now, and stick with this line of dialogue. What sort of new capabilities do you’ve gotten coming on-line within the close to future, and when? What needs to be fascinated by by way of finest utilizing your strengths?
Claude: Straight reply first: I can’t see Anthropic’s inner roadmap, so I received’t hand you a dated record of unreleased options — that’s exactly the confabulation entice I warned you about, and it’d be worse coming from me proper after that warning. What I can do is inform you what’s truly shipped within the hole between my coaching cutoff and at this time, as a result of quite a lot of it’s instantly related to you and also you will not be utilizing it but. A few of it reads much less like “close to future” and extra like “already right here, you simply haven’t turned it on.” 2
Now the sincere caveat that ought to form how you utilize all of this. Even the finance-tuned setup scored about 64% on a finance agent benchmark — a failure charge that, as one author put it, would get a human analyst tossed, and Anthropic’s personal framing is that customers ought to keep firmly within the loop, reviewing and approving Claude’s work earlier than it goes to a consumer or will get filed. So the trajectory isn’t “Claude replaces the verification step.” It’s “Claude does extra of the meeting, and your judgment turns into the scarce, helpful enter.” 3
The work that will get extra helpful is the editorial judgment, the thesis, the voice, and the fact-checking — the components which might be distinctly Barry Ritholtz and may’t be benchmarked.
So reasonably than chasing every characteristic, I’d place for the pattern: get the recurring workflows specified tightly sufficient that they survive being handed to a extra autonomous agent, and preserve your hand firmly on the verification gate. The tooling will preserve getting higher on the first half by itself.
Barry: OK, remaining query – this one is about Co-Work and Code:
I had
you[Perplexity] create an interactive Drake Equation dashboard – it was very cool, however it didn’t depend on altering information. I attempted another dashboard concepts, ran take a look at runs; it was not fairly prepared for prime time.What are you able to create that auto-updates based mostly on (weekly, month-to-month, quarterly) financial and (every day, weekly, month-to-month) market information?
I need to create financial & market dashboards that present varied information collection at a look; that get up to date robotically; I need to embed it onto WordPress so anybody [with that link] can see it; and I need to set particular indicators that create alerts based mostly on particular (private) parameters that I set…
Claude: That is the suitable query, and the reply explains why your Drake dashboard and the opposite take a look at runs felt caught. It wasn’t a polish drawback — it was structural. An artifact (the factor I generate in chat) is a self-contained front-end: no backend, no scheduler, no secure place to retailer an API key, and solely a flaky potential to name outdoors information sources from the browser. So it might probably look nice however it has nothing feeding it and nothing waking it as much as refresh.
What you’re describing isn’t a fancier artifact; it’s a small information pipeline plus a hosted web page. The excellent news is that Claude Code and Cowork are constructed to create precisely that — the artifact was simply the unsuitable container.
Right here’s the structure that does every part in your record, and the one design resolution that issues most for you.
Break up it into two halves: a public dashboard and a personal alert engine — fed by the identical pipeline.
The pipeline is a scheduled job that runs on a cron (say, each weekday after market shut, plus a month-to-month run for the gradual financial collection). Every run it pulls contemporary information, computes your indicators, writes out a small JSON (JavaScript Object Notation) snapshot, and checks your alert thresholds.
The important thing transfer: your private parameters stay inside this job, server-side, the place nobody can see them. The general public dashboard by no means accommodates your thresholds — it simply reads the pre-computed JSON snapshot and attracts charts. So anybody embedding or viewing the WordPress web page sees the info at a look, whereas your alert logic and secret ranges keep fully personal and fireplace solely to you. That separation is the factor most individuals get unsuitable: if the thresholds had been within the front-end, “view supply” would expose them immediately.
Mapping that to your 4 necessities:
Auto-updating comes from the cron, not the web page. Financial collection (CPI, unemployment, yield curve, and so on.) refresh on their very own month-to-month/quarterly cadence; market collection refresh every day. The dashboard is quick as a result of the heavy lifting already occurred — it simply masses a baked JSON.
Knowledge sources. FRED (the St. Louis Fed’s API) is the gold commonplace for financial collection — free, deep, and importantly, redistributable, which issues because you’re publishing publicly. Market information is the half to watch out with: free daily-close sources exist, however real-time change information carries licensing restrictions on public redistribution. For a public embed I’d persist with every day closes and clearly-licensed collection; that sidesteps a headache you don’t need as a registered agency. (Value having Code confirm the present phrases of no matter market supplier you decide, reasonably than trusting my recall on free-tier limits.)
WordPress embed is the straightforward half: the dashboard will get hosted as a static web page someplace, and also you drop an iframe (or Customized HTML block) in your WordPress web page pointing at it. It renders for everybody, and it updates at any time when the cron republishes the JSON — you by no means contact WordPress once more.
Alerts fireplace from the scheduled job to wherever you need them — e mail, SMS, a Slack/Discord webhook. Your parameters keep within the job’s secret config.
The sincere half about what I can and may’t do. Code can write all of this: the fetch scripts, your indicator math, the alert engine, the cron configuration, and the front-end charts — and take a look at it in its sandbox. What it might probably’t do is be the server that runs eternally. The scheduled job has to stay someplace. On your state of affairs the cleanest path that requires no server to babysit is: a GitHub repo + GitHub Actions (free cron) working the pipeline, the JSON snapshot printed to a free static host (GitHub Pages or Cloudflare Pages), API keys saved as encrypted Actions secrets and techniques, and alerts despatched by way of an e mail/webhook service. You personal three issues — the data-provider accounts, the place it’s hosted, and the key thresholds — and Code builds and wires up every part else.
That’s an actual, finishable challenge, not a sometime factor. A wise first construct is a working vertical slice: decide three or 4 FRED collection and one market collection, get up the pipeline + one chart + one alert finish to finish, affirm it refreshes and embeds, then develop the collection record and indicators as soon as the plumbing is confirmed.
One resolution drives the stack: do you need to run this completely on free hosted infrastructure you principally by no means log into (the GitHub Actions + Pages route above), or do you have already got a server / Vercel / Cloudflare setup you’d reasonably use? Inform me which, and I’ll have Code scaffold the slice — pipeline, chart, alert, and the WordPress iframe snippet — so you possibly can see it replace by itself earlier than we construct it out.
`
Footnotes:
1. Claude: Quantitative scaffolding and constructing issues. The interactive React viz work — variance drain, the Marks paradox — performs to an actual power: I can flip an idea right into a working, calibrated visible artifact in a single go.
2. Claude: “The opposite shift value understanding: the present mannequin, Opus 4.8, was tuned in a route that’s related to my largest weak spot for you. Anthropic says early testers report it’s extra prone to flag uncertainties about its work and fewer prone to make unsupported claims, and it added consumer management over how a lot effort Claude places right into a activity. That doesn’t remove confabulation, however it’s shifting the suitable manner — and the trouble management is beneficial while you desire a quick reformat versus a cautious evaluation.”
3. Claude: “That’s the true reply to “what ought to I be fascinated by.” The route of journey is clearly towards agentic, tool-connected work that lives contained in the apps you already use and runs longer chains autonomously. What that rewards is unglamorous and sturdy: clear, well-organized inputs; your own home conventions written down as soon as so any agent applies them constantly; and a verification self-discipline you belief. The work that will get automated is the meeting and formatting you’re already handing me.”
