A working explanation, not a sales page. What the thing is, how it runs, what it produces, and the five things only you can supply. If a section is unclear, say so and it gets rewritten.
You asked how this works before deciding anything. Fair. Here it is without jargon.
An agent, the way we build it, is not a chat window that tells you what it would do. It runs on its own machine with a terminal, a browser, a file system and access to the open internet. It writes code and runs it, reads pages and documents, and comes back with finished work. The difference is the one between someone describing how to build a landing page and someone handing you the landing page.
For you the practical version is narrower and more useful: it takes over the search and reconnaissance part of prospecting, the part that eats the week, and leaves you the part you said you wanted to keep, which is judgment and the conversation.
It finds and qualifies. You decide and approach. Nothing is sent to a human being without a person saying yes to the exact words first.
Not a mock-up or a slide. This is the output of a working version, run on public sources only, from the signal definition you gave: hiring for Growth Marketing, Performance Marketing, Paid Acquisition, User Acquisition and Demand Gen, filtered toward SaaS and AI, mobile, established DTC and ecommerce, in Canada and the United States.
Measured on our own run on 26 September, from public job boards, with no paid data tool involved. The files behind them, the raw postings and the list itself, are available to you if you want to check them rather than take our word for it.
Each surviving company came back with the company and website, the hiring signal itself, why it fits an offer like yours, what to lead with, the likely decision-maker, and an opening line tied to the specific posting. The reasoning is attached to the row, so you can check the thinking instead of trusting it.
Those postings are public. Companies publish their own job boards as open data. That means the discovery layer does not depend on scraping LinkedIn, on a data subscription, or on anyone's API key. It also means it does not break when a platform changes its terms.
Four parts, and they explain why the output gets sharper instead of staler.
The part that thinks. A setting, not a commitment: it can be swapped and nothing else about the system changes.
The part that acts. A terminal, a browser, files, the open web, and a long list of smaller ones behind those.
Notes and context that survive between sessions. Your rules and your corrections stay in place.
Procedures it writes for itself the first time it does something, and reuses afterwards. The second run is cheaper than the first.
Five steps, in this order. Every one of them is auditable, which matters when you are deciding whether to trust a row.
Three differences, and they are the reason the output gets better rather than staler.
Who counts as a fit, what disqualifies a company, how much each signal is worth: all of it sits in a short list at the top of the system, in plain text. Changing a rule does not mean rebuilding anything.
When you say that a company was a bad call, or that the second tier was actually the interesting one, that correction stays for the next run. The system is meant to accumulate your judgment, not repeat your first draft of it.
The first time it does something awkward, it keeps the working method. The tenth run of that task is faster and more consistent than the first, without anyone documenting the steps by hand.
Prospecting does not fail because the method is wrong. It fails because a busy week arrives and nobody runs it. The result is a pipeline that quietly empties while the work still looks done.
Running this on a clock changes the failure mode. The list arrives whether or not anyone remembered, and the value compounds because the search keeps watching companies you would never have checked again.
Four deliveries you did not have to ask for, each one comparable against the last, so you can see which companies got louder (more roles, new channels) and which went quiet.
Straight answer, and it is the important part of this document. It arrives knowing nothing about Crux Media. It cannot guess what a good prospect is for you, and it will produce confident-looking nonsense if nobody tells it.
| It needs from you | Why the system cannot invent it |
|---|---|
| Your ideal client, in your own words | Fit is a judgment about who you want to work with, not a property of a company. It can read size and market. It cannot read whether you want that client. |
| What you sell, at what price, and what an engagement looks like | This decides which signal matters. A retainer business and a project business read the same hiring post differently. |
| Two or three cases worth showing | The pitch angle and the opening line are grounded in something real. Without a case, the line becomes generic. |
| What makes you rule a company out early | Your disqualifiers are worth more than your wish list. They cut the list faster than any scoring tweak. |
| How many prospects a week makes this worth it | Without a number there is no way to say whether it is working. Five good ones and fifty mediocre ones are different products. |
No logins. No ad account access. No API keys. No paid data subscription. Your numbers stay on your side, and the disclosure goes the other way: you tell us about your business, not about your accounts.
Three things this does not do, stated plainly so there are no surprises later.
Two things you could do instead, and an honest read on both.
| Approach | What it gives you | Where it hurts |
|---|---|---|
| Hiring a junior for prospecting | Judgment in the loop from day one | Weeks of onboarding, a salary that runs whether or not the list gets built, and one person's capacity as the ceiling |
| A data subscription | Enormous raw databases you can query | It is data, not conclusions. The filter, the reasoning and the opening line are still your job, and you are back to doing the work weekly |
| This | A shortlist with reasoning, arriving on a schedule | Needs your input to be worth anything, and it should be judged after four or five cycles, not one |
The five inputs in the table above, in whatever form is easiest: a voice note, a rough document, a list of bullet points. From those we write the specification and show you the first version.
So you know exactly what you are getting. Every shortlisted company arrives with the same set of fields, in a spreadsheet and in a machine-readable file.
| Field | What it carries |
|---|---|
| Company and website | Who they are and where to look |
| Signal | The posting itself, with a direct link, so nothing rests on a summary |
| Why they fit | Two or three sentences of reasoning, checkable against the posting |
| What to lead with | The angle, tied to the channel the posting says they are building |
| Who to approach | The likely decision-maker for that spend |
| Opening line | One sentence grounded in the posting, ready to edit rather than invent |
| Score and tier | How the weighting judged it, so you can disagree with the weighting |
The score exists to sort the list, not to be believed. If a row scores high and the reasoning reads wrong, the rule is wrong, and that is a five-minute fix that improves every run after it.
You also asked about producing short-form creative at a workable cost. It is the same machine with a different job, so it is worth putting the numbers in front of you now.
The creative figures are our own costing, built from published tool pricing and the rates agencies quote publicly. They are a planning range, not a quote, and if your own numbers differ we would rather hear it than defend ours.
Two things stand out. Generation itself is nearly free, cents per image and under three dollars per clip. The cost sits in the human loop: the brief, the selection, and the check that the claim in the ad is true. Second, an accepted ad is not a generated asset. A static gets a name, a size set and a landing page match. A video gets a script, cuts, sound and captions.
That gives you a number to plan against: one person with the system in place covers roughly 40 to 80 statics or 10 to 20 finished videos in a week. At the rates above, a monthly arrangement of around 30 statics and 10 videos lands in the 3,000 to 6,000 dollar band, which sits inside the range you said your clients can pay.
Generation is a commodity and its price falls every month. Anyone can buy the same access. The defensible part is the loop around it: testing against winners, packaging and naming that keeps ad platforms happy, a hard quality gate that stops generated junk from going live, and a claim check, since regulated verticals switch an ad off over wording rather than image quality.
No phased proposal, no retainer discussion before it works. What actually happens if you hand over the inputs.
Three numbers, agreed up front, so this does not become a matter of opinion.
If after four cycles the replies are not there, the input definition is wrong or the approach is wrong for your market, and we say so rather than selling you a fifth month. Stating that now is easier than arguing about it later.
It runs on its own machine, reads public sources, applies rules you own, and produces a shortlist with reasoning on a schedule. It is not a chatbot and not a database. The first run read about five thousand postings and returned twenty-one companies worth looking at, six of them strong, without a paid tool or a key. What it needs is your definition of a good client, what you sell, a couple of cases, what makes you say no, and a weekly number that would make this worth it.
One list, in one place, so it does not get lost in the reading. In any form you like: a voice note, a rough document, a few bullet points.
Send those and we write the specification, then show you the first version. That is the whole next step.
Written to be read once and understood. If anything is unclear, ask and it gets rewritten in the same plain register.
Yulyana
ARLI membership: the notes from the workshop
One site of yours, measured, with the price for the work in the same document
ARLI, Inc., a Delaware corporation. Questions: hq@arliwork.com.