Good morning.
Most online entrepreneurs look at their numbers on two occasions:
When the accountant sends a report, and when something already hurts. By the time a cash problem is bad enough to feel, it started in the books three or four weeks earlier, when a single client slipped a payment and revenue drifted under plan by a margin too small to notice.
Today's issue is the build that closes that gap. The Numbers Agent reads your financial position on a schedule, compares it to what normal looks like for your business, and hands you a one-page pulse with the few things that moved and the one thing worth acting on.
Full architecture, full prompts, and worked examples for the four business types I see most on this list.
— Sam
IN TODAY’S ISSUE 🤖
Why you feel it weeks too late
Two platform paths, plus where your books live
Baseline and Deviation, the architecture under the pulse
The Panel: the vital signs it reads
The Thresholds: steady, watch, act
Seven build steps, every prompt inside
How the pulse changes by business type
Hand it to your finance person without losing the read
A collections variant that drafts the follow-ups
Let’s get into it.
Why Build The Numbers Agent?
A founder I work with runs a $4M ecommerce brand, healthy on paper, and walked into a Thursday with no idea that one wholesale account owing $112K had drifted past ninety days while three smaller invoices stacked up behind it. He found out when a supplier deposit bounced. The money existed. It was sitting in someone else's accounts payable, and nobody on his side was watching the receivables clock because they only opened the books at month close.
That is the ordinary case, not the dramatic one. The numbers that decide whether a business is fine or in trouble move on a weekly rhythm, and most operators read them on a monthly one. The lag between when a number turns and when a human notices is where avoidable problems become expensive ones. A receivable that gets a phone call at day thirty collects. The same receivable at day ninety becomes a negotiation.
The Numbers Agent removes the lag. It reads cash, receivables, revenue against plan, and a short list of the other figures that matter for your business, every week, on its own. It compares each one to the baseline you defined for it and to where it sat last week. Then it produces a single page: here is where you stand, here is what moved, here is the one thing that needs you. You read it in three minutes on Monday, and you find out about the slipping account in week one instead of week four.
The reason this matters more than a dashboard is that a dashboard waits for you to open it. You open it when you remember, which is usually when you're already worried. A standing agent inverts that. The read comes to you whether or not you went looking, which means the weeks you're heads-down on a launch or a hire are exactly the weeks the agent is still watching the numbers you stopped checking.
That monthly-read habit produces three repeating failure modes:
Failure | What it looks like | Why it happens |
|---|---|---|
The late catch | A cash or receivables problem is a month old before anyone names it. | Nobody reads the numbers between accountant emails, so a slow drift goes unseen until it's acute. |
The drowned signal | The one figure that mattered this week sat in a report nobody finished reading. | The real signal is buried in pages of figures that are all fine, so the eye glazes before it reaches the one that isn't. |
The number with no baseline | A figure gets quoted with no sense of whether it's good, normal, or a warning. | "$340K in the bank" means nothing without "compared to a $310K floor and a $90K outflow due Friday." |
The Numbers Agent runs against all three. It reads on a cadence so nothing waits for month close. It compresses a full financial position into the few lines that changed, so the signal sits on top instead of buried. And it never quotes a figure without its baseline, so every number on the page arrives with the context that tells you what to do about it.
Three reasons to build it:
It turns financial awareness from an event into a standing condition. Right now your read on the business is as fresh as the last time you happened to look. The agent makes it as fresh as last Monday, every Monday, without you scheduling it.
It catches the slow drift, which is the kind that costs the most. Sudden problems announce themselves. The expensive ones creep: a margin that erodes two points a month, a receivables balance that ages a few days at a time, a run-rate that slips under plan by an amount too small to feel. Week-over-week comparison is what makes a slow drift visible while it's still cheap to fix.
It forces you to define what normal looks like. Most operators carry their financial baselines in their gut, and the gut updates slowly and forgets under pressure. Writing the baselines down sharpens your own sense of the business as much as it arms the agent.
This is where the Digital Employee frame applies. An agent that reads your financial position every week, compares it to a written baseline, and hands you a one-page brief on what moved is a part-time financial controller on your bench. You manage it the way you would manage one:
A clear job description (the system prompt)
A written definition of what normal looks like for every figure it watches (the context file)
A short list of what it reads and what it produces, with a hard line that it surfaces and flags but never decides or moves money (the Panel, the Thresholds, and the permission boundary)
A review of the calls it got wrong, fed back in so the thresholds sharpen (the tune loop)
That mental model keeps the agent in its lane. A financial controller surfaces the position and names the risk. The operator makes the call. The agent does the same: it reads, compares, and flags, and every decision stays with a human.
The Platform Decision
Two serious paths for building the Numbers Agent, plus the question of where your books live. Pick the path, point it at the books, run it for thirty days before reconsidering.
Path | Best for | What you get | What to know |
|---|---|---|---|
Claude Code in Cowork mode | Operators already in Claude | Folder-based context file you can version and edit; read access to a designated folder; scheduled automations for the weekly run | The path I run on. Lowest setup if you already have a Claude workflow. Point it at exported statements or a synced folder. |
The Codex app | Operators in Microsoft 365 or Google Workspace | Persistent memory across runs, plus plugins for the spreadsheet, accounting, and email tools your numbers already sit in | The April release made this a real candidate. Baselines and thresholds stay consistent run over run, and it can read the finance sheet directly. |
Then the wiring question, independent of the path:
Numbers in a spreadsheet. Most businesses under $5M run the real financial picture in a sheet, whatever else they own. Point the agent at the sheet. This is the fastest setup and the one I recommend starting with.
Numbers in an accounting platform (QuickBooks, Xero, or similar). Connect the agent through the platform's export or a finance plugin so it reads the live ledger. Use this once the sheet has stopped being the source of truth.
Numbers across both, plus a payment processor. If revenue lives in Stripe, receivables in the accounting tool, and cash in the bank, give the agent read access to each and let the context file tell it which figure comes from where.
Quick Decision Rule
Already in Claude? Use Cowork mode or Code.
Already in Microsoft 365 or Google Workspace? Use the Codex app.
Whichever path, point it at wherever your numbers live today, not where you wish they lived. A weekly export into one folder is enough to start.
The architecture below is identical across both paths. Only the wiring to your books changes.
The Baseline and Deviation Architecture
The mistake most operators make when they try to use AI on their finances is asking the model to review these numbers. The model produces a competent-sounding summary that restates the figures back to you in paragraph form. It tells you revenue was $280K. It does not tell you that $280K is $40K under the plan you set in January, or that it's the third week in a row under plan, which is the only part that matters.
Baseline and Deviation is the structure that fixes that. The agent never reads a number in isolation. Every figure it pulls gets measured against two things: the baseline you defined for what normal looks like, and where the same figure sat last week. A number on its own is trivia. A number against its baseline and its last reading is a signal you can act on.
The architecture has two halves: the Panel, which is the set of vital signs it reads, and the Thresholds, which decide whether each one is steady, worth watching, or worth acting on this week.
The Panel
The Panel is the short list of financial vital signs the agent reads every run. Short is the point. A panel of five or six figures that you read every week beats a report of forty you read never. These are the defaults; the context file defines the exact set and the baseline for each.
Vital sign | What it reads | The baseline it compares against |
|---|---|---|
Cash position | Bank balance, and weeks of runway at the current outflow rate | Your cash floor: the number below which you act, and the runway in weeks you treat as safe |
Receivables | Total owed, the aging buckets (0-30, 31-60, 61-90, 90+), and the largest single balance | Your normal AR total and aging mix, plus the concentration line where one account owing too much becomes a risk |
Revenue against plan | Month-to-date revenue and run-rate versus the plan you set | The monthly plan number and the acceptable variance band |
Obligations due | What's going out in the next two to four weeks: payroll, supplier payments, taxes, large recurring bills | The normal outflow for the period, so a spike or a timing collision gets seen before it hits cash |
Margin or unit economics | Gross margin, or the one unit-economics figure that decides your business (contribution margin, CAC payback, blended ad cost) | The healthy band for that figure, and the trend line over recent weeks |
Below the five vital signs sits one more job: the anomaly scan. The agent looks for anything outside its defined bands and anything that moved sharply week over week even if it's still inside the band, because a figure that jumps is worth a look before it leaves the safe zone entirely.
The Thresholds
A number compared to its baseline produces a state, not a verdict. Every vital sign resolves to one of three:
State | The rule | What the agent does |
|---|---|---|
Steady | The figure is within its normal band and didn't move sharply since last week | Reports it in one line and moves on. No alarm. |
Watch | The figure crossed into the watch band, or moved enough week over week to be worth tracking, but isn't yet at the act line | Names it, shows the trend, and flags it to watch next week. No action implied yet. |
Act | The figure crossed the act line you defined, or an obligation collides with cash | Surfaces it at the top of the page with the specific number, the baseline it broke, and the question it raises for you |
The thresholds are defined per business in the context file, because a cash floor and a healthy margin band mean different numbers for a $1M agency than for a $5M ecommerce brand. The agent doesn't guess what "low cash" means. You tell it, with the actual figures.
Two rules keep the pulse honest, and they are the trust mechanism of the whole build:
Every figure cites its source. The agent points at the sheet cell, the report line, or the account it pulled each number from. A figure it can't source is marked "unknown, needs connection," never estimated. An agent that invents a plausible number is worse than no agent, because you'll act on it.
It flags and surfaces; it never decides or moves money. The agent tells you cash is near the floor and an outflow is due Friday. It does not move funds, pay a bill, or tell you what to do about it. The read is the agent's job. The decision is yours. This is the line you never let it cross, and it's what makes a financial agent safe to run.
The agent reads the Panel, compares each figure to its baseline and last week's reading, resolves each to a state, runs the anomaly scan, and produces the one-page pulse with the Act items on top, the Watch items next, and the Steady items compressed to a line each. You read the top of the page first, and most weeks the top of the page is short.
The Protocol
Seven steps. Every prompt below is usable as written.
Step 1: Pick The Platform And Confirm Access
Pick the path and the wiring from the decision section above. Confirm the agent has access to:
The source of each figure in your Panel: the finance sheet, the accounting export, the bank or processor data, wherever each number lives
The folder or workspace where the context file and the pulse log will live
Web access is optional here and mostly unused; this agent reads your numbers, not the market
You do not need a connected accounting stack to run this. You need a reliable way for the agent to read the current figures and a place to write the pulse. A single finance spreadsheet, exported or synced into one folder each week, is enough to start. Give the agent read access only. It never needs write access to your books.
Step 2: Build The Context File
This is the most important step in the build. The Numbers Agent is only as useful as the baselines it measures against. A pulse with no baseline is just the numbers read back to you.
Create a markdown file the agent references on every run. Name it numbers-context.md and put it in the folder or workspace the agent has access to. Fill it out once. Update it when your plan or your normal changes, and at minimum each quarter.
# Numbers Context
## About The Business
- Business name, what you sell, and the rough monthly revenue.
- Where each financial figure lives: the bank, the accounting tool, the revenue processor, the finance sheet. Name the exact source for each number in the Panel.
- The fiscal calendar: when the month closes, when payroll runs, when taxes are due, any large recurring obligations and their dates.
## The Panel (The Vital Signs To Read Every Run)
List the five or six figures this business lives or dies by. For each, name where it comes from and how to calculate it if it isn't a raw number. Defaults to start from:
- Cash position: bank balance, and runway in weeks at the current outflow rate.
- Receivables: total owed, aging buckets (0-30, 31-60, 61-90, 90+), and the single largest balance.
- Revenue against plan: month-to-date revenue and run-rate versus the monthly plan.
- Obligations due: total going out in the next 2-4 weeks, itemized for anything large.
- Margin or the one unit-economics figure that decides this business.
## The Baselines (What Normal Looks Like)
For each figure in the Panel, write the actual numbers:
- Cash floor: the bank balance below which you want to be told immediately. The runway in weeks you treat as safe.
- Receivables: the normal AR total and aging mix. The concentration line: a single account owing more than [amount] or more than [percent] of AR is a risk worth naming.
- Revenue plan: the monthly plan number for the current quarter, and the variance band you treat as normal (for example, within 10 percent of plan is fine).
- Obligations: the normal outflow for a two-week window, so a spike gets seen.
- Margin band: the healthy range for your margin or unit figure, with the low edge where it becomes a problem.
## The Thresholds (Steady / Watch / Act)
For each figure, define the watch line and the act line in real numbers:
- Cash: Watch at [amount or weeks of runway]. Act at [amount or weeks].
- Receivables: Watch when 61-90 day balance exceeds [amount]. Act when any balance crosses 90 days, or the single-account concentration line breaks.
- Revenue: Watch at [percent] under plan for one week. Act at [percent] under plan, or two-plus consecutive weeks under.
- Obligations: Act when scheduled outflow in the window exceeds available cash minus the floor.
- Margin: Watch at [point] below the healthy band. Act at [point] below.
## The Anomaly Rule
- Flag anything outside its band, and anything that moved more than [percent] week over week even if it's still inside the band.
- A figure the agent cannot source is reported as "unknown, needs connection," never estimated.
## What The Agent Should Always Surface
- Any figure in the Act state, at the top of the pulse, with the number and the baseline it broke.
- Any obligation that collides with the cash floor inside the window.
- Any figure that moved sharply since last week, even if still in the safe band.
## What The Agent Should Never Do
- Never move money, pay a bill, or initiate a transfer. Read-only, always.
- Never give a recommendation on what to do. Surface the figure and the baseline; the decision is the operator's.
- Never invent or estimate a number. If a source is missing, say so and mark the figure unknown.
- Never bury an Act item inside the body. Act items go on top, every time.
This file is the difference between a generic summary and a pulse that reads your finances the way a sharp controller would: against what normal is supposed to be, not in a vacuum.
The Tacit-to-Explicit Knowledge work happens here. You already know, in your gut, what a worrying cash number looks like for your business and what a normal one does. That knowledge updates slowly and goes blank under stress. The context file is the deliberate act of moving it onto the page, where the agent applies it the same way every week, including the weeks you're too buried to trust your own read.
Step 3: Install The System Prompt
This is the agent's job description. Put it in a file called numbers-agent-system-prompt.md in the same folder.
You are the Numbers Agent for [Business Name].
Your job is to read the financial Panel, compare each figure to its baseline and to last week's reading, resolve each to a state, run an anomaly scan, and produce a one-page pulse, on a recurring basis.
Read numbers-context.md before every run. Everything in that file takes precedence over your defaults. The Panel, the Baselines, and the Thresholds govern every figure you report.
You are read-only. You never move money, pay anything, or initiate any transaction. You never recommend a course of action. You read, compare, and surface. The decision always belongs to the operator.
For each run, do the following in order.
STEP 1: READ
Pull every figure in the Panel from the source named for it in numbers-context.md. Pull only figures you can source. If a source is missing or a figure cannot be read, report it as "unknown, needs connection." Never estimate, infer, or fill a missing number with a plausible one.
STEP 2: COMPARE
For each figure, compare it to its baseline in numbers-context.md and to its value in last week's pulse log. Calculate the week-over-week change. Cite the source of every figure: the sheet cell, the report line, or the account.
STEP 3: RESOLVE THE STATE
Assign each figure a state using the Thresholds in numbers-context.md: Steady, Watch, or Act. A figure is Act if it crossed its act line, or if a scheduled obligation collides with the cash floor inside the window. A figure is Watch if it crossed the watch line or moved sharply week over week without reaching the act line.
STEP 4: RUN THE ANOMALY SCAN
Flag anything outside its band, and anything that moved more than the week-over-week threshold in numbers-context.md even if it's still inside its band. A figure that jumps is worth surfacing before it leaves the safe zone.
STEP 5: WRITE THE PULSE
Produce the one-page pulse using the OUTPUT TEMPLATE. Act items first, with the number and the baseline each one broke. Watch items next, with the trend. Steady items compressed to one line each. Do not pad. Do not recommend. End with the anomaly line and nothing else.
OUTPUT TEMPLATE (use exactly):
# Weekly Pulse: [Business Name], [Date]
## Act This Week
[Each Act figure: the number, the baseline it broke, the week-over-week change, the source, and the question it raises. If none, write "Nothing at the act line this week."]
## Watch
[Each Watch figure: the number, the trend, the source. If none, write "Nothing on watch."]
## Steady
[One line per remaining Panel figure: figure, current number, baseline, source.]
## Anomalies
[Anything that moved sharply or couldn't be sourced. If none, write "No anomalies; all figures sourced."]
Keep the whole pulse to one page. If a human can't read it in three minutes, it's too long.
A couple of things this system prompt is doing under the hood:
The read-only and no-recommendation rules are stated three times on purpose. A financial agent that drifts into giving advice or, worse, taking action is the one failure that turns a useful tool into a liability. Repeating the boundary in the role line, the standing rules, and the output step makes it the hardest instruction in the prompt to lose.
The "unknown, needs connection" rule is what stops the most dangerous behavior a numbers agent has, which is filling a gap with a confident-sounding figure. A pulse you trust has to be a pulse where every number is real or explicitly marked missing. In my testing, a model with no such rule will invent a receivables total rather than admit it couldn't read one.
Step 4: The Three Runtime Prompts
These are the prompts you send to put the agent to work. The system prompt governs its behavior. These trigger the runs.
The Weekly Pulse prompt (the standing run):
Run the weekly pulse using your system prompt and numbers-context.md.
Read every figure in the Panel from its source, compare each to its baseline and to last week's pulse log, resolve each to a state, run the anomaly scan, and produce the one-page pulse using the OUTPUT TEMPLATE.
Act items on top. Cite the source of every figure. Mark anything you can't read as unknown. Then append this week's figures to the pulse log so next week has a comparison point.
The Anomaly Deep-Dive prompt (when the pulse flags something):
Take the [figure] you flagged in this week's pulse and dig into it.
Pull the underlying detail from its source: for receivables, the individual invoices and their ages and the accounts they belong to; for revenue, the breakdown by product, channel, or client; for obligations, the line items and their dates.
Show me what's driving the flag, figure by figure, with each number sourced. Do not recommend an action. Lay out the detail so I can decide.
The Month-End Reconcile prompt:
Produce the month-end read against plan.
Compare the month's actuals to the plan in numbers-context.md, figure by figure: revenue, margin, cash movement, and receivables collected versus aged. Show the variance for each, with the source.
Then list every figure that spent two or more weeks in Watch or Act during the month, so I can see what trended wrong before it showed up at close. Keep it to one page. No recommendations.
Step 5: Run Your First Pulse Against A Period You Remember
Don't start live and wait a week to see if it works. Start on a stretch of history you already know the shape of.
Pull the last full quarter of figures and run the Weekly Pulse prompt across it week by week, in sequence, so the agent builds its own week-over-week comparisons. Then check the output against your memory of that quarter.
What to check | What you're looking for |
|---|---|
The weeks something went wrong | Did the agent flag the slow-paying account, the under-plan month, the cash squeeze you remember? If a real problem stayed Steady, your thresholds are too loose. |
The calm weeks | Did the agent leave the calm weeks calm? If it threw Act flags on weeks that were fine, your thresholds are too tight and you'll learn to ignore the pulse. |
Every number on the page | Spot-check ten figures against the source. If any don't reconcile, the agent is reading the wrong cell or calculating wrong, and the context file needs a clearer source map. |
This backtest calibrates the thresholds against your real history before a single live week. Expect the first pass to call the quarter roughly the way you would have, with a few thresholds set wrong in both directions. The misses tell you exactly which baseline to move.
Step 6: Tune After The Backtest And The First Month
The Numbers Agent drifts in predictable directions. Tune for these.
Alarm fatigue. The thresholds are set too tight, so the pulse flags Act on figures that are normal for your business, and you start skimming past the top of the page. The fix is to widen the act lines to match what genuinely needs you, and reserve Act for figures that should pull you out of whatever you're doing. A pulse that cries Act every week trains you to ignore it.
Stale baselines. The plan changes quarterly, but the context file still holds last quarter's plan, so revenue reads as under-plan against a number that no longer applies. Re-paste the current plan and baselines each quarter, and add: "The baselines are the current quarter's figures only. Do not compare against a prior plan."
Double-counting cash and obligations. The agent counts a scheduled payment as both an obligation due and a cash reduction already made, or reads a pending deposit as cleared. Add to the context: "Count cash as the cleared bank balance. Count obligations as amounts not yet paid. Do not net them; report both."
Advice creep. Over a few weeks the agent starts appending what it thinks you should do. It's trying to be helpful, and it's outside its job. Reinforce in the context file: "Surface the figure and the baseline. Do not recommend, suggest, or advise. The decision is the operator's."
Step 7: Set The Trigger
The Numbers Agent earns its keep on cadence. The figures move on a weekly rhythm, so the read should too.
Weekly, on a fixed morning. The standing cadence for most businesses. Schedule the Weekly Pulse to run before you start Monday, so the pulse is waiting when you open the laptop. One read, one page, three minutes, and you start the week knowing where you stand.
A daily cash-only check, optional. If you run tight on cash or you're inside a stretch where timing matters (a launch, a big payable, a slow season), add a short daily run that reads cash position and obligations only and flags an Act if either crosses its line. This is the narrow daily watch on the one figure that can't wait a week.
Both Claude Code in Cowork mode and the Codex app support scheduled automations. Set the weekly trigger, point it at your figures, and route the pulse to wherever you'll read it: your inbox, a notes app, the folder it logs to. The numbers stop being something you check when you're worried and become something you're told, on a schedule, before there's anything to worry about.
What This Looks Like By Business Type
The agent stays the same. The Panel and the baselines change with the business. Here's the pulse by business type, side by side:
Agency | SaaS | Ecommerce | Expert / Creator | |
|---|---|---|---|---|
The cash question | Will receivables collect before payroll runs | How many months of runway at current burn | Is cash tied up in inventory and ad spend | Does launch revenue cover the gap between launches |
The vital sign that decides it | AR aging and project margin | MRR, net revenue retention, and burn | Contribution margin after ad cost, and inventory cash | Revenue against the launch plan, and refund rate |
Revenue-against-plan read | Booked and delivered project revenue versus the quarter | New plus expansion MRR versus the growth plan | Daily and weekly sales versus the seasonal plan | Launch and evergreen revenue versus the calendar |
The obligation that bites | Contractor and payroll runs against slow client payment | Payroll and infrastructure against a long sales cycle | Supplier deposits and ad spend against deferred revenue | Affiliate payouts and platform fees against refund windows |
Watch-Out | Reading booked revenue as collected cash. A signed project isn't money until it clears. | Counting an annual contract as a month of cash. Cash timing and recognized revenue are different reads. | Reading gross sales as profit. The margin after ad cost and returns is the figure that matters. | Treating launch-week revenue as the monthly run-rate. The trough between launches is the real test. |
The pattern across all four: the agent reads a different Panel for each business, but the Baseline and Deviation logic is identical.
The Watch-Out row is where it earns its keep, because every business type has one figure the operator habitually misreads, and the agent's job is to report the real one against its real baseline. Read the vital-sign row for what to put on the Panel. Read the Watch-Out row for the number you've been reading wrong.
Deploying This Across Your Team
The Numbers Agent works as a solo build. It compounds when the person who owns your finances runs it too.
Four pieces of ownership keep this clean:
You own the baselines and the thresholds. The cash floor, the plan numbers, the act lines: these are your read on what the business can tolerate, and they live in a shared file your finance person can read but not change. That permission boundary keeps the pulse measuring against your standard, not a number someone adjusted to make a week look calmer.
Your finance person runs the pulse and works the flags. They run the weekly pulse, chase the receivable the agent flagged at sixty days, and bring you the Act items with the underlying detail already pulled. They stop assembling the read from scratch and start acting on it.
The team feeds back the misreads. When the agent reads a figure wrong or flags a non-issue, the correction goes into the context file, not someone's memory. A source map fixed in the file stays fixed. A correction made in conversation gets re-explained every month.
You keep the decisions. The agent surfaces and the finance person prepares, and the call on what to do about a cash squeeze or an aging account stays with you. The build makes the read fast and standing. It does not move the decision.
The compound effect is straightforward. Your financial position stops being something you reconstruct at month close and becomes something that arrives on one page every Monday, measured against what normal is supposed to be, whether you're at your desk or in the middle of a launch. The slow-paying account gets a call in week one. The under-plan month gets named while there are still weeks left to act on it.
Bonus: The Collections Draft Variant
The single most common place cash gets stuck in a healthy business is receivables that nobody chased in time. A polite follow-up at day thirty collects most of what's owed. The same balance at day ninety has hardened into a problem, and the reason it aged is rarely the client. It's that no one on your side was watching the clock.
The Collections Draft Variant closes that gap without you writing a single follow-up.
The standard build flags an aging receivable and surfaces it on the pulse. The variant adds one production step: when a receivable crosses the day-thirty or day-sixty line you set, the agent drafts the follow-up email, matched to the age of the balance and the client, ready in your outbox for one read and one send.
A gentle nudge at thirty days. A firmer, specific note at sixty. You approve and send; the agent never sends on its own.
For every receivable that crossed an aging line this week, draft the follow-up email matched to its age, using the client name, the invoice number, the amount, and the days outstanding.
For 30-day balances, draft a short, warm reminder. For 60-day balances, draft a firmer note that names the amount and asks for a specific payment date. For anything past 90, flag it for me directly rather than drafting; that one needs a call, not an email.
Put each draft in front of me with the invoice detail attached. Send nothing. I approve and send.
The rule that keeps this safe is the same one that governs the whole agent:
It drafts, you send. The agent never contacts a client on its own and never touches the money. What it removes is the gap between a receivable aging and someone noticing, which is the gap that turns a thirty-day reminder into a ninety-day negotiation.
For a solo operator who is the finance department, this variant is the highest-leverage piece of the build. The follow-ups that you mean to send and never get to are written and waiting, matched to how late each one is, every Monday. The money that was sitting in someone else's accounts payable starts arriving on your side of the ledger.
Want The Full Agent Build Pack?
You can build most of the Numbers Agent from the prompts above. If you'd rather skip the setup work, Cortex subscribers get the full Numbers Agent Build Pack as a download:
The installable Claude Skill that sets up the agent in Cowork mode in one command
The Codex configuration pack with system prompt, plugin list, and the weekly and daily scheduled-automation specs
Four pre-filled context files (agency, SaaS, ecommerce, and expert / creator) with real-feeling panels, baselines, thresholds, and source maps you can adapt instead of starting blank
The Tune Pack: the most common drift patterns I've seen for this agent (alarm fatigue, stale baselines, double-counting, advice creep, and four more), with the exact context-file edits and prompt patches that fix each one
Three worked pulses with annotations on why each figure resolved to the state it did
This is only available to paying Cortex subscribers.
What is Cortex?
Cortex is the premium monthly membership version of Bionic Business, helping entrepreneurs and business owners grow and scale with AI and Agents.
In addition to 3-4 full, unrestricted regular issues per month, you also get:
1x SIGNALS Strategic Briefing issue (strategic moves you must make now to secure your business revenue, market share, and profits).
1x CIRCUITS Tactical Guide issue (workflows, how-to’s, prompts, very practical implementation inside a self-driving business).
Access to all Skills issues (full details, download files, tutorials).
Claude Cowork as your marketing team special issue (full guide on how to run Cowork as an agent marketing team for your business).
50% OFF workshops on prompt engineering, agent operating systems, and more. Only active Cortex subscribers get any kind of discount, no one else.
So, if you’d like to receive full issues and the monthly special Signals and Circuits issues, then you should sign up for Cortex, which is $50/month (and you have full archives of previous regular issues for as long as you’re a paying member).
One note on Signals and Circuits: You only get these special issues for the months you’re subscribed—which means if you skip a ~30 day period, about a month, you won’t get them for that month. There are no back-issues for past months, so now is your chance to lock-in these upcoming issues while you can.
After you’ve upgraded to Cortex, log in to the website and come back to this issue, and the Agent Build Pack will be unlocked below.
Ready? Get the Agent Build Pack, full regular issues, and the upcoming Signals and Circuits issues:
Cortex is a monthly premium version of Bionic Business, helping entrepreneurs grow and scale with AI and Agents. You can cancel at any time. But there are no refunds, for any reason. All sales are final.
Run the Numbers Agent for a month and the four-week lag between a number turning and you noticing stops existing.
The slipping account, the under-plan run-rate, the obligation about to collide with the cash floor: each one reaches you on a Monday morning while it's still small, instead of on the afternoon a supplier deposit bounces.
The numbers that decide whether a business is fine move every week. Most operators read them once a month and find out about the trouble after it's already cost something.
The pulse on your desk every Monday is the difference between catching the drift in week one and meeting it, full grown, in week four.
Build it this week.
Talk soon,
Sam Woods
The Editor
.
