Simply Broken
Exploring / NoonFox: CMO Agent / v1
Exploration · 5 September 2026

The CMO you cannot hire

A pre-launch iPhone app with one person behind it needs a chief marketing officer and can afford exactly none. So seven agents built one. Then an eighth was told to tear it apart, and it found five things a generic CMO would have shipped on day one.

7
agents
22
critique verdicts
5
must-fix
1
north star
$0
first 30 days
01 · The brief

A generic CMO is a horoscope

The ask was specific. Not a chief marketing officer, but the chief marketing officer this product needs at this exact moment: proactive, data-driven, technically deep, elegant, frugal. The failure mode was obvious from the start. Ask any model for a CMO persona and you get something that reads well, flatters the reader, and would describe a dating app, a bank and a garden centre equally well.

So the test applied to every sentence in the finished persona was blunt. Would this line read the same way for any other product? If yes, it is not a principle, it is decoration, and it was cut. Six lines failed that test in the first draft and were named individually by the critic.

The product is a photo-first home inventory app. You snap a photo, the fox remembers what is in it, and later you ask where something is in plain words.

It is called NoonFox. Native iPhone, one price of $3.95 a month per person locked forever, household sharing included where each member pays their own way. Nineteen customer types have been researched and scored: the person with ADHD who cannot find anything, the serial mover with forty unlabelled boxes, the collector, the person documenting a house for insurance. It has no notification system, on purpose, which turns out to matter enormously to how it can be marketed.

The constraint that shapes everything: the marketing budget is not money, it is one founder's attention.

Nothing sends itself. Every outbound message gets read and pressed by a human, which means a plan that calls for two hundred emails is not a cheap plan, it is an impossible one. So every week in the finished plan carries two costs: dollars, which are zero, and founder minutes, which are counted and capped.

02 · The build

Four researchers, one editor, one adversary

The persona was not written in one pass. It was orchestrated: four researchers working in parallel on four questions that do not overlap, then a consolidation, then an adversarial critique whose only job was to refute, then a revision that had to answer every finding.

Diagram: four parallel research agents feed one consolidation agent, which feeds an adversarial critic, which feeds a final revision. The product, as it really is How the best growth leads think Frugal iPhone playbooks, 2026 Analytics and attribution codebase, store metadata, vault their own essays and talks platform docs, dated benchmarks what is knowable about a stranger refute, do not improve ConsolidateCritiqueRevise one draft22 verdictsthe persona
The four researchers never see each other's work, which is the point. The critic never sees the researchers' instructions, only their evidence and the draft, so it judges the draft against the facts rather than against the intent.

The critic was told to refute, not to improve. That single word changes what comes back.

An agent asked to improve a draft returns polish: tighter sentences, a better heading, one more bullet. An agent asked to refute a draft returns the sentence that is false, the number that is unsourced, and the plan step that cannot physically happen. It returned twenty-two findings, listed the six lines that were generic, and named twelve things that were missing. Every finding had to be answered before the persona shipped.

03 · The interesting part

Five things a generic CMO would have shipped

These are the findings that justify the whole exercise. Each one reads perfectly well in a marketing plan. Each one is wrong for this product, and only wrong because of a specific fact the critic went and checked.

Must-fix 01

Outreach scheduled at an app nobody can install

The draft had week three answering ten forum threads and writing to twenty creators. Reasonable, energetic, and the standard advice.

The app is not in the store yet. Its listing returns a not-found page. Every one of those messages would have pointed at nothing, burned the one first impression each of those communities gives you, and taught us nothing measurable. Week three now drafts all thirty messages and sends none. They go out the week after approval, with a real link attached.

Must-fix 02

An offer the product cannot honour

Creators were to be offered a free year of the app.

There is no billing code in the app at all. No purchase can be made, so no year can be given. The offer became early beta access now, and a real store-issued code for twelve months once billing ships. The disclosure requirement came with it: a gifted product has to be declared by the creator, and any offer that reads as payment for a review is a policy violation that gets developers removed from the store.

Must-fix 03

The wrong first customer, chosen for the right-sounding reason

Lead with the person documenting their home for insurance. Hurricane season is live, the need is acute, the searches exist.

That job ends in a document, and the app cannot export one yet. A competitor in that niche already exports to a spreadsheet. Marketing a promise the product cannot keep is the most expensive mistake available, because it converts and then churns. The first customer became the person mid-move looking for a specific box, whose whole job is search, which does work today. Insurance waits for an export feature and a January timing.

Must-fix 04

A differentiation claim that was false in our own files

Every competitor needs manual entry or barcode scanning, so we are the only one that recognises items from a photo.

Our own customer research names a competitor that does exactly that. The line had survived because nobody re-read the file it came from. The claim was rewritten to what actually holds up, and a standing rule was added: never write the word "only" until a dated re-check of the three named competitors is on file.

Must-fix 05

A day-thirty scorecard full of numbers that cannot exist

By day thirty we will know which customer type converts better and which channel pays back.

The app is submitted for review in week four. Nothing that requires a live store listing can have a value thirty days from a standing start. The section was rewritten to what is genuinely knowable: the store state, one reconciled set of listing copy, the activation rate of about twenty to forty testers with an honest note that the sample is too small to kill anything on, and whether the measurement shipped at all.

The pattern under all five: each was a plan step that assumed a capability the product does not have. Store listing, billing, export, a defensible claim, a live funnel. A marketing plan is a set of promises about what someone else can do next, and every one of those promises has a dependency in the codebase.

None of the five would have been caught by writing more carefully. They were caught by sending something to check the repository, the store and the competitor file, and by asking it to disprove rather than to help.

04 · The persona

Seven principles, and what each one refuses

What survived is short. A principle earns its place by forbidding something specific that somebody would otherwise do this week.

No outreach wave until the retention curve flattens. Growth before that point only increases churn.

This is the one point on which every growth writer of the last decade agrees, and it is the one most often skipped. A churned user cannot be sold to later. The key action here is not the photo, it is the successful find, so the curve that has to flatten is the find curve.

Measure usage before spending anything on reach. Money spent before measurement exists teaches nothing.

Today the app ships no analytics of any kind. That is a deliberate privacy posture, not an oversight, and it means the first marketing job is not a campaign, it is instrumentation that stays inside the privacy promise: anonymous, not linked to a person, not tracking anyone across other apps.

One who, one channel, one lever, one week. Nineteen researched customer types are a menu, not a plan.

The persona picks the customer on Monday from the numbers and does not change it mid-week. Running three customer types across two channels in one week produces a result that cannot be attributed to anything, which is the same as no result at a much higher cost in attention.

Little channels, never paid. At this price the arithmetic is not close.

Published 2025 benchmarks put the cost of one installed utility app on the platform's own ad network at about $2.90. Published 2026 subscription data puts the first-year value of a paying customer for a low-priced app at about $10.69, at a download-to-paying rate whose median is around two percent. That is roughly $145 of advertising to buy one paying customer worth about eleven dollars. Advertising here is a research tool with a $100 credit, not a channel.

Feed the one loop that actually exists. Everything else is a cost, not a compounding asset.

Household sharing is a paid invite loop: one member invites another, and the second member pays their own subscription. That is a weaker loop than a free invite and it should be expected to perform at the low end, which is exactly why it has to be measured in three separate steps rather than counted as a feature.

Privacy is the positioning, not a constraint on it. No advertising identifier, no tracking prompt, no ad network.

This is the line that decides the analytics choice, rules out an entire class of advertising, and makes attribution harder. It is not traded away for a better dashboard. An app that holds photographs of the inside of your house has exactly one thing it cannot afford to be casual about.

Founder minutes are the budget line. Every week states its cost in minutes beside its cost in dollars.

This was the critic's finding, not the draft's idea. The plan said cost zero four times and never counted the fifty individual approvals it was quietly asking for. A plan that is free in money and unaffordable in attention is not a cheap plan.

05 · The scoreboard

One number, and the seven that explain it

North star
Weekly found users

People who opened a search result this week. It is the brand promise stated as a measurement: the fox remembered, and the person believed it. Installs cannot inflate it and a demo cannot fake it. The weekly cadence is an estimate for now, to be re-checked against eight weeks of real search data rather than asserted forever.

Every threshold below is paired with a minimum sample. No number moves a decision until thirty people are behind it.

Another critique finding. The first draft had kill rules that needed eight weeks of cohort data and store benchmarks that do not exist yet for a product with no listing. A rule you cannot execute is worse than no rule, because it looks like rigour. Below the sample floor the instruction is explicit and unglamorous: look at the funnel and talk to ten people.

Captures in the first seven days
The input that can move before the north star does. If people are not capturing, nothing downstream can happen.
TurnA median under five sends the week to the onboarding work, not to marketing.
Activation rate
Captured something, then opened a search result. In the first session, and again within a week. This is the moment the product becomes true.
TurnUnder half of new users in-session stops all outreach.
Day-28 retention of found users
Measured on the people who have actually found something, not on everyone who downloaded.
TurnTwo cohorts under four percent means the product is the problem, not the channel.
Trial to paid
Reads zero until billing exists. Published data says most cancellations of a short trial happen on the first day, which makes the first session the whole game.
TurnUnder fifteen percent reopens the trial length question.
Conversion per custom store page
The cheapest per-customer-type experiment available. One page per audience, each reporting its own conversion.
TurnUnder half the default page, after 200 impressions, gets one refresh then retirement.
Downloads per campaign link
The only honest attribution left on this platform. A token on the link, counted by the store itself.
TurnA channel under the reporting floor after three weeks is closed.
Household invites: sent, accepted, paid
Three separate numbers, because the drop between them is where the loop actually lives.
TurnPaid invites under five percent means it is a feature, not a channel.

What it refuses to report

Website traffic. Total downloads. Followers. Impressions. Press mentions. Beta build counts. Email signups presented as a result rather than as a denominator. Each of these can be made to go up by someone trying, without a single additional person finding their own belongings, which is the definition of the problem.

06 · The reading

What it took from the field, and what it left

The research read the primary sources rather than the summaries of them. The useful output was not a reading list, it was a shortlist of famous ideas that are correct in general and wrong here, with the reason attached.

Kept

Retention before acquisition
Plot the curve of the key action at its natural frequency, act when it flattens.
Channel and model fit
A low-priced product fits word of mouth, community and search. The middle of the price range is where products die.
The smallest viable audience
Choose who it is for, then go where they already are. Reach almost nobody, deliberately.
Activation before retention
Churn is usually a symptom of a first session that never delivered the point.
One lever per weekly sprint
Find the tenth of the work that carries most of the result, and stop the rest.
Every channel decays
So favour the small channels the large competitors ignore, and accept wins that do not scale.

Left on the shelf

Streaks and leaderboards
Built for a daily-habit product. A home inventory is not opened daily and pretending otherwise designs the wrong loop.
Performance marketing at launch
Needs cohort analytics and five-figure monthly spend to learn anything. The arithmetic above closes it.
Atomic networks and the cold start
Beautiful theory, and this product has no cross-user network effect to bootstrap.
The full four-fits exercise
A framing for a company at scale. One line of it applies here; the rest is a later question.
Playing every square of the growth matrix
Written for business software with a sales team. There is one lever that matters this month.
Store page A/B tests
The platform needs volume this product will not have for months. Below it, a test is theatre with a confidence interval.
07 · The spine

How it says no

A persona that only proposes is a brochure. These are the five refusals it is expected to make, in the words it makes them in.

"Not before the curve."
To buying installs before retention is proven. The arithmetic says every one of them loses money, and a churned user cannot be recovered later.
"Say the find, not the feature."
To any headline that leads with the technology. The buyer has a question, which is where their thing is. Naming the model that answers it is not an answer.
"No token, no channel."
To any effort on a channel that cannot be measured back to a download. An unmeasured win teaches nothing and cannot be repeated on purpose.
"Not a promise the app cannot keep."
To any audience whose job ends in something the product does not do yet. This is the one that moved the first customer.
"Featuring is a reward, not a plan."
To any growth model containing the phrase "and then the store features us". Nominate early, then plan as though it will not happen.
08 · The recommendation

Thirty days, zero dollars

The plan the persona opened with. Every week names the number it moves and the minutes it costs, and every step that needs a human decision is marked as one.

Week one · about 20 founder minutes

Build the ruler before measuring anything

  • Verify the real store state through the store's own interface rather than from a document written in May.
  • Write the weekly numbers script and run it once by hand, against the database, the crash reporter and the issue tracker.
  • Open the analytics change: fifteen events, one activation definition, and the privacy declaration in the same change so they cannot drift apart.
  • Reconcile two contradicting drafts of the store listing into one that matches the price the product actually charges.
Moves nothing yet. This is the week that makes the rest measurable.
Week two · about 30 founder minutes

The store page is the product, for most people

  • Name, subtitle and keywords. The description written in the plain words people search, because the store now generates its own tags from that metadata.
  • Six screenshots pulled from real recorded app footage, the first two showing snap then find, short benefit captions, no superlatives.
  • Twenty external testers recruited from the existing outreach database, screened for the operating system version the app requires.
Moves activation rate, from nothing to a first reading.
Week three · 60 founder minutes, capped

Prepare the first audience, send nothing

  • One custom store page for the first customer type, with its own keywords assigned.
  • Ten forum answers drafted with the link left as a placeholder. Nothing posts until the listing exists.
  • Twenty creator notes drafted, offering early access now and a real code later, asking for honesty and a disclosure and nothing else.
Moves activation and captures per new user, on a cohort of twenty to forty.
Week four · about 30 founder minutes

Submit, quietly

  • Submit for review. This is a human decision, not an automated one.
  • Ask for a review after the first successful find, never on launch.
  • Index items into system search, which is the way back into an app that deliberately has no notifications.
  • Stage the website switch from coming soon to a real store link carrying a campaign token, ready for the day it goes live.
Moves crash-free rate on the release build. Everything else waits for approval.

What day thirty actually tells us. The verified store state and one clean set of listing copy. The activation rate of twenty to forty testers, with the honest caveat that the sample is too small to kill anything and is there to guide the first session, not the marketing. The crash rate. And whether the measurement shipped at all, which decides whether the north star reads anything in month two.

Not: conversion, channel payback, or trial economics. Those need a live listing, and saying so on day zero is the difference between a plan and a wish.

09 · The honest part

What is still open

The persona ships with six questions it cannot answer for itself, because each is a decision about the product rather than about its marketing. Whether the first release carries billing at all. Whether the trial stays at three days. Which forum account has enough history to be trusted. Whether the mover really is the right first audience. Whether the analytics change ships in week one or waits. Whether export gets built in time for the insurance audience in January.

Each of the six changes a specific line of the plan, and the plan states the assumption it is running on until the answer arrives. A persona that pretends to have no dependencies is the same failure as a plan that assumes capabilities the product does not have. It just hides better.

The reusable lesson, if there is one. Send something to check, and tell it to disprove.

Four researchers produced a persona that read extremely well and contained five plan steps that could not physically execute. One adversary with access to the same evidence and a mandate to refute found all five in a single pass. The cost of that pass was a few minutes. The cost of not running it would have been a launch month spent messaging people about an app they could not download.

NoonFox is a photo-first home inventory app for iPhone. This page documents how its marketing persona was built and what the critique changed, not the product itself. All versions of this topic · More explorations