25:35Jack Roberts · Tutorial
Every Level of Hermes Agent Explained
A 25-minute level-by-level breakdown of Hermes agent, from one-shot prompts to a model-agnostic agentic OS that ships work while you sleep.
June 17thThe argument in one line.
Framing Claude as an analyst rather than a dashboard builder — and leaving the interface unspecified — is the single decision that determines whether the output explains what drove the numbers or merely displays them.
Read if. Skip if.
READ IF YOU ARE…
- You work in FP&A, treasury, or ops finance and spend time manually reconciling spreadsheets, ERP exports, and email threads into one picture.
- You want a reporting environment that explains the operational drivers behind every metric, not just the metric itself.
- You already use Google Sheets for financial actuals and want analysis to live on top of those same files without a data migration.
- You are evaluating Claude Cowork live artifacts and want to see a production-grade use case beyond a one-off chat output.
SKIP IF…
- You need enterprise audit trails, role-based access controls, or SOC 2 compliance for the reporting environment.
- Your financial data lives in systems Claude Cowork cannot connect to — SAP, Oracle ERP, Snowflake, proprietary data warehouses.
The full version, fast.
Claude Cowork live artifacts stay connected to external sources and refresh as those sources change. This tutorial builds a seven-layer FP&A command center for a fictional automotive group using Google Drive as the financial backbone, Airtable as the planning and scenario layer, and Gmail as a live operational signal feed. The core prompt design insight is to frame Claude as an analyst rather than a dashboard — which shifts output from chart generation to cross-source reasoning that links supplier delays and overtime spikes to their EBITDA impact. The resulting environment updates when source spreadsheets are edited directly, with no Zapier, Make.com, or automation middleware required.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Create a free account →Where the time goes.
00:00 – 01:08
01 · Cold open
Hook: the system reads financial reports, operational data, treasury metrics, workforce data, and business emails to explain what is happening and why.
01:08 – 02:57
02 · What makes live artifacts different
Contrast: normal Claude artifact is a snapshot; live artifact inside Cowork stays connected to external sources and refreshes as data changes. Multi-source reasoning during the build phase.
02:57 – 06:27
03 · Connecting the sources
Three-source architecture: Drive (financial backbone), Airtable (planning and scenario layer), Gmail (operational signal layer). Four key prompt design decisions explained.
06:27 – 08:56
04 · How Claude builds the environment
Discovery phase before interface generation: reads all sources, builds cross-source relationships, parses exact numeric values via bash, validates calculations before registering as persistent artifact.
08:56 – 14:50
05 · Walking through the command center
Seven layers: Executive Pulse, Variance Intelligence, Revenue & Margin Drivers, Plants & Supply Chain, Workforce & Cost, Cash & Treasury, Operational Signals (~$95M EBITDA risk from Gmail threads).
14:50 – 17:10
06 · Modifying the source data
Live demo: edits Google Sheets directly, refreshes artifact twice, values flow through. No rebuild, no prompt, no automation infrastructure.
Lines worth screenshotting.
- Telling Claude to act as an FP&A analyst instead of building a dashboard shifts its output from chart generation to causal reasoning about what drove the numbers.
- Prescribing the interface layout produces a hard-coded demo; leaving layout unspecified produces an environment that fits the actual business data.
- Assigning a distinct role to each data source — backbone, assumption layer, signal feed — prevents Claude from treating all sources as interchangeable and losing cross-source reasoning.
- Claude Cowork live artifacts do not need Zapier, Make.com, or any scheduled job; the artifact itself is the connection layer between source files and the analysis.
- A discovery phase before interface generation — reading sources, building relationships, validating numbers — is what separates an FP&A system from a demo dashboard.
- Parsing exact numeric values from spreadsheets via bash during the build keeps the analysis mathematically grounded rather than interpretive.
- Operational email signals (supplier delays, overtime escalations) can be translated into quantified forecast assumption revisions before they ever reach a reporting cycle.
- Three active Gmail signal threads translated into roughly $95M of first-half EBITDA risk — categorized as managed vs. unmanaged exposures.
- A revenue beat driven by volume has different strategic implications than one driven by pricing execution; a reasoning environment makes that distinction visible automatically.
- The finance team keeps working in the same Google Sheets they already use; the live artifact moves with their edits without requiring them to change their workflow.
- Covenant headroom monitoring is more valuable than covenant reporting — knowing where leverage peaked tells treasury how much operational deterioration the balance sheet can absorb.
- Live artifacts require two refresh cycles when source data changes because the synchronization layer processes the updated state in the background before new values flow through.
How framing changes what Claude builds.
WHAT TO LEARN
The most important decision in this build is not which data sources to connect — it is whether you ask Claude for a dashboard or an analyst, because that framing determines whether the output generates charts or generates reasoning.
- Telling Claude to behave as an FP&A analyst rather than build a dashboard shifts what it optimizes for — causal reasoning about what drove the numbers replaces chart generation.
- Leaving the interface unspecified produces a more useful environment than prescribing layout, because Claude designs around the actual business data rather than a wireframe imagined in advance.
- Assigning each data source a distinct role — backbone, assumption layer, signal feed — is what allows cross-source reasoning to work; without it, Claude treats all sources as interchangeable.
- The discovery phase before interface generation mirrors how a real analyst works: reconcile the data first, validate the calculations, then construct the view — not the other way around.
- Operational email signals (supplier delays, overtime escalations, pricing pressure) can be translated into quantified forecast revisions before they reach a reporting cycle, turning inbox noise into financial model inputs.
- Live artifacts refresh from source file edits, not from prompts — finance teams keep working in the same Google Sheets and the analysis layer moves with them automatically.
- No automation middleware is required: the live artifact is the connection layer between source files and the analysis environment, replacing Zapier flows and scheduled database jobs.
Terms worth knowing.
- Live artifact
- A Claude Cowork output that stays connected to external data sources and refreshes its values and charts as those sources change, unlike a standard artifact which is a static snapshot from a single conversation.
- FP&A
- Financial Planning and Analysis — the finance function responsible for budgeting, forecasting, variance analysis, and connecting operational performance to financial outcomes.
- Claude Cowork
- A Claude feature that allows connecting external data sources (Google Drive, Airtable, Gmail, etc.) to a session so Claude can reason across live business data during artifact generation.
- Revenue bridge
- A decomposition of the variance between actual and budgeted revenue into its component drivers — volume, pricing, mix, FX, regional — showing how much each factor contributed to the overall gap.
- EBITDA exposure
- The estimated financial impact on earnings before interest, taxes, depreciation, and amortization from a specific operational risk or event, expressed as a dollar amount.
- Covenant headroom
- The gap between a company's current leverage ratio and the maximum ratio permitted by its debt agreements — a measure of how much financial deterioration is possible before a debt covenant is breached.
- Operational signals layer
- In this system, the Gmail-connected section that reads operational threads and translates them into quantified adjustments to forward-looking forecast assumptions.
Things they pointed at.
Lines you could clip.
04:44
“The difference between an environment that displays financial data and one that explains it usually comes down to this framing.”
05:52
“If you describe the layout, the charts, the workflows, and every component in detail, you get a hard coded demo back basically.”
07:12
“Claude does not summarize these sources separately. It starts building relationships between them.”
14:05
“Three active operational signals together represent roughly $95,000,000 of full year twenty six first half EBITDA risk.”
16:05
“There is no Zapier flow. There is no make.com automation. There is no cloud database. The live artifact itself is the connection layer.”
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
analogy
00:00I built a live AI FP and A analyst inside Claude live artifacts that continuously reads financial reports, operational data, treasury metrics, manufacturing performance, workforce planning, and even business emails to automatically explain what's happening inside the business and why.
00:20So instead of behaving like a static dashboard, the system stays connected to live business data, links operational activity directly into financial impact, identifies risks before they hit reporting cycles, and behaves more like a real finance intelligence environment than traditional reporting.
00:38The entire system runs across connected business sources in real time, so the analysis evolves as the underlying operational and financial data changes. As a quant and finance automation developer, I'm gonna show you exactly how I built this unified FP and A environment inside Claude Live Artifacts.
00:58And if you want the prompts and the full project setup from this video, it's free in the community I have linked in the description below. So for now, let's get into it. Before we build anything, I wanna explain what live artifacts actually are because that distinction is what makes this project possible.
01:17A normal Claude artifact is something most people have seen. You ask Claude for a dashboard or a document inside a conversation, and then Claude generates it, and the file sits beside the chat.
01:29And you can iterate on it during that session, but it essentially it's a snapshot. It doesn't stay connected to anything and it does not update when the underlying data changes. A live artifact inside Cowork is built differently.
01:45It can stay connected to external sources. So things like a Google Drive folder or an air table base and Gmail, things like that.
01:53Now the metrics and the charts inside it then refresh when the source data changes. So the artifact reflects the current state of the business rather than just a frozen view that doesn't change.
02:06So the second thing that matters is what Claude can pull together when it builds a live artifact.
02:15Normal Claude outputs work from whatever is in the conversation at that moment. With a live artifact, Claude can reason across multiple connected sources during the build, structured financial data, operational planning assumptions, communication, and then bake that analysis into the environment itself.
02:36The multi source reasoning is what gives the system its analytical depth from the very start. Now once it is built, the data layer stays connected and the values keep refreshing as the sources change.
02:50The analytical commentary written into the artifact reflects the picture Claude built it from. Connecting the sources.
02:58I opened a new session inside Claude Cowork and confirmed that three sources are already connected. The Google Drive folder named as an example use case. Aurora Automotive Group holds the financial and operational backbone, things like monthly actuals, budgets, treasury reports, manufacturing data, supply records, and executive commentary.
03:21The Airtable base holds the structured planning and the operational assumptions, including risk flags, scenario inputs, and KPI targets. And the Gmail account holds the operational signal layer where supplier delays, overtime escalations, pricing pressure, and production updates show up as live business messages.
03:42Each source plays a different role inside the system. Drive is the historical and current financial reality.
03:51Airtable is the forward looking assumption layer. Gmail is the live operational signal layer that basically flows business events into financial implications before those events ever reach a reporting cycle.
04:05Now, the prompt. There are a few decisions inside of that shape what the system becomes, and they are worth understanding before you build something like this yourself.
04:15The first decision is to frame the artifact as an FP and A analyst rather than a dashboard.
04:22Now that single framing shifts what the system prioritizes. So instead of defaulting toward chart generation, Claude shifts toward reasoning.
04:33It thinks about what drove the numbers, what operational conditions created the movement, and what risks exist underneath the surface. And the difference between an environment that displays financial data and one that explains it usually comes down to this framing.
04:52The second decision is to ask for one cohesive environment rather than separate mini tools. Real FP and A analysis is interconnected. Various analysis connects to cash flow, operational issues connect to margins, planning assumptions, influence forecasting.
05:11Now if those layers are all just fragmented into disconnected widgets, then the system can't reason across them. One unified environment is what allows the intelligence to work across the connected sources all at once.
05:25The third decision is to assign a clear purpose to each connected source.
05:32So Claude knows which system is the financial backbone, which is the assumption layer, and which is the operational signal feed. Now that is what stops the system from treating them as interchangeable.
05:46The fourth decision is not to prescribe the interface.
05:50If you describe the layout, the charts, the workflows, and every component in detail, you get a hard coded demo back basically. Letting Claude design the interface itself, guided by the analytical role and the data available, produces something that actually fits the business content it was built for.
06:09Now the prompt also tells the system to test calculations, filters and live behaviors before presenting the artifact, which pushes Claude into a validation mindset rather than a generation only mindset when it is connecting to multiple life sources at once. How Chlord builds the environment.
06:28Once the prompt runs, Chlord does not start generating an interface. It moves into a discovery phase first where it reads through every connector source and tries to understand the business environment before building anything on top of it.
06:41It identifies the drive folder as the financial and operational backbone. It pulls the actuals and budget files. It opens a supplier and dealer datasets and then reads the treasury reports.
06:52Then it moves into air table to extract the assumption layer, the scenario records, and the operational risk mapping. Then scans Gmail and pulls operational threads tied to overtime, supplier delays, pricing pressure, production disruptions.
07:09Now what stands out here is that Claude does not summarize these sources separately. It starts building relationships between them.
07:17It is connecting how operational events in Gmail might affect financial outcomes in the drive folder, or how planning assumptions in Airtable should shape forecast interpretation, and where contextual business activity ties back into reported numbers.
07:34Now very early in the process, it explicitly states that it needs to retrieve specific numeric values from the actuals and budget files by pausing them through bash. Now that is significant moment because it shows the system moving past surface level interpretation.
07:53It is putting exact financial values directly so that the analysis stays mathematically grounded. The build then becomes increasingly data driven.
08:03Claude reconciles revenue bridges, validates gross profit calculations, links operational events to EBITDA exposure, and checks whether scenario assumptions match planning logic.
08:15It tests interactive navigation filters and chart relationships before finalizing anything. And only after the full environment behaves consistently does it register system as a persistent live artifact inside co work.
08:31This sequence matters because it is how a real FP and A analyst would approach the same problem. Before building reporting layers, an analyst reconciles source data, validates assumptions, identifies operational drivers, and connects business events to financial outcomes.
08:50So Claw just follows the same workflow during the build itself. Walking through the command center. The final environment is a connected intelligence system with seven analytical layers.
09:02Each one focused on a different operational domain, but all tied to the same financial backbone. The first layer is the executive pulse. This is the leadership facing command view that summarizes overall financial health while explaining the operational conditions behind the numbers.
09:20At the top, Claude surfaces revenue, gross margin, EBITDA, net income, and net leverage against budget.
09:28And what makes this layer different from a normal KPI dashboard is that every metric is paired with the operational reasoning behind it. The revenue beat is tied to volume growth, pricing realization, and EV mix expansion.
09:43The margin improvement is tied to lower battery costs, freight normalization, and premium mix. The leverage peak in August is tied to treasury pressure even and a recommendation to pre clear debt funded k p x until covenant headroom improves.
10:03Every KPI carries its own driver story. The various intelligence layer is built like a real bridge analysis.
10:12You can switch between revenue, EBITDA, COGS, and net income views, and then compare actuals against either budget or the q three forecast. The revenue bridge decomposes the favorable variance into specific operational drivers including volume recovery, EV mix shifts, pricing realization, FX headwinds, and regional softness.
10:37Each driver has its own commentary explaining why the movement happened and whether it is structural or just temporary. The monthly p and l underneath shows how performance evolved through the year, and the regional and product line breakdown separate consolidated results into the segments that actually drive them.
10:56The revenue and margin driver section goes a bit deeper into the mechanics behind the financial results. So revenue growth is broken into its components. The distinction between what drove revenue and what drove margin is what makes the analysis actionable rather than just descriptive.
11:13A revenue beat driven by volume has different strategic implications than one driven by pricing execution or product mix. And the system makes that visible rather than just rolling everything into a single headline number.
11:27The manufacturing and supply chain layer tracks production operations across multiple plants. Utilization, output volumes, downtime events, supply delays, and logistics costs are all present.
11:42And what matters about this section is how it connects into the financial environment rather than operating as just a standalone operational report. A supply chain disruption does not just appear as an operational flag, it flows through to margin pressure and it shows up in the relevant financial sections as well.
12:01The operational and financial layers also inform each other throughout the system. The workforce section connects labor conditions into cost structure and profitability, Overtime levels, attrition, hiring gaps, and wage inflation are mapped against manufacturing throughput and EBITDA exposure.
12:21The system goes beyond tracking the metrics. When overtime runs significantly above budget across a set of plants, it identifies the affected facilities, it estimates the financial impact, and then it links it back to the margin picture. And that connection between workforce execution and financial outcome is something finance teams usually have to calculate and document separately.
12:45And here, it's just all built in. The treasury section tracks liquidity, leverage ratios, covenant headroom, debt structure, and capital deployment across the fiscal year.
12:56Now one of the most valuable things this section does is surface the peak leverage point during the year alongside current headroom.
13:06So that historical context matters for covenant monitoring.
13:11Knowing what leverage came close to the covenant cap during a specific period tells treasury leadership where the sensitivity is and how much operational deterioration the balance sheet can absorb before financing risk increases.
13:26And that kind of forward looking awareness is what separates treasury monitoring from treasury reporting. The operational signals layer is the most advanced part of the system.
13:37So instead of waiting for month end reporting, it pulls signals directly from Gmail and translates them into forecast implications. And over time, escalation across three plants becomes an upward revision to labor cost growth assumptions.
13:52A North American pricing pressure thread becomes an increase in dealer incentive assumptions. A semiconductor delay in one of the European plants becomes a reduction in plant utilization assumptions for the coming quarter. Three active operational signals together represent roughly $95,000,000 of full year twenty six first half EBITDA risk.
14:15And the system already separates managed exposures from unmanaged ones.
14:21Now what makes this layer matter is the operational or is the fact that operational events stop sitting in someone's inbox, and they start moving directly into forecast assumptions.
14:34So supplier delays, overtime spikes, pricing pressure, and production disruptions become quantified inputs to the financial model rather than just disconnected updates that someone has to manually translate later on.
14:49Modifying the source data. This is the part that proves the system is actually connected rather than presenting fixed values.
14:57So I opened the full year 2025 monthly financial actuals Google Sheet directly, the same file that artifacts is reading from, and I changed the July and December revenue values inside the consolidated p and l section.
15:13The EBITDA, operating income, tax expense and net income values update automatically because the sheets formulas carry through. Then I go back to the live artifact and click refresh.
15:25The first refresh does not pick up the changes because the synchronization layer is still processing the updated state in the background. So I wait a few moments and then I click refresh again and the new values now flow through.
15:39The KPI cards update the revenue trend chart redraws with the new July and December spikes. The artifact is responding to your change I made inside the source file not to a prompt.
15:52I did not rebuild the dashboard. I did not regenerate any visuals. I just edited a Google Sheet and the FP and A environment moved with it.
16:01Now what makes this powerful is that no automation infrastructure is involved.
16:07There is no Zapier flow. There's no make.com automation.
16:12There's no cloud database. There's no scheduled job pulling values into into the dashboard. The live artifact itself is the connection layer.
16:22The finance team continues working inside the same Google Sheets. They already use and the executive environment evolves alongside that work.
16:32So that's it. That's the full system. One prompt, three connected sources, and then seven analytical layers and a live artifact that updates as the underlying business data updates.
16:43The Aurora command center is doing the work that normally lives across spreadsheets, ERP exports, planning models, treasury schedules, manufacturing trackers, and email threads.
16:56All of it collapsed into one continuously operating environment. Now, if you want the prompts and the full project setup from this video, it's in my free community which is linked in the description below.
17:09And I'll see you next time.
The Hook
The bait, then the rug-pull.
What if your financial reporting environment explained the numbers instead of just displaying them — and updated itself whenever you edited a spreadsheet? That is the premise of this tutorial, which walks through building a seven-layer FP&A command center inside Claude Cowork that stays connected to Google Drive, Airtable, and Gmail simultaneously, with no automation middleware required.
Named ideas worth stealing.
04:07list
Four Prompt Design Decisions
- Frame Claude as an analyst, not a dashboard
- Ask for one cohesive environment, not separate mini-tools
- Assign a clear purpose to each connected source
- Do not prescribe the interface
The four decisions inside the prompt that determine whether the output is a reasoning environment or a hard-coded demo.
Steal forany Claude Cowork project where you want cross-source reasoning rather than a chart generator
02:42model
Three-Layer Source Architecture
- Drive = historical and current financial reality (backbone)
- Airtable = forward-looking assumption layer (planning, scenarios, risk flags)
- Gmail = live operational signal layer (events flow into financial implications before month-end)
Assigning distinct roles to each source type prevents Claude from treating them as interchangeable and enables layered cross-source reasoning.
Steal forany multi-source AI agent where data has different temporal horizons (historical vs. forward-looking vs. real-time)
08:56list
Seven-Layer FP&A Command Center
- Executive Pulse
- Variance Intelligence
- Revenue & Margin Drivers
- Plants & Supply Chain
- Workforce & Cost
- Cash & Treasury
- Operational Signals
The seven domains covered by the unified command center, each connected to the same financial backbone.
Steal forstructuring any finance AI environment where operational and financial data need to inform each other
How they asked for the click.
VERBAL ASK
16:45link
“If you want the prompts and the full project setup from this video, it is in my free community which is linked in the description below.”
Soft, repeated twice (once at ~1:00 and once at close). Links to community platform. Clean, low-pressure, consistent with educational creator positioning.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
