THIS FORM IS THE PRODUCT CONFIG. EVERY FIELD BECOMES A LINE THE AGENT RUNS.
Onboarding a client here is supposed to be a directory plus a warmed sending
domain. If it needs code, the abstraction leaked and that is our bug, not your problem. So this
form is not a lead-capture form pretending to be a conversation — it is the campaign pack, typed in by
you, field for field. Scroll and you will watch the YAML build itself at the bottom of the page.
It takes about 25 minutes and you cannot fake your way through the proof-pack section.
That is deliberate. Read the part in red before you start typing.
// WHAT THIS AGENT HAS ACTUALLY DONE, AS OF 2026-08-31
Before you fill in 30 fields, here is the whole record. Nothing is rounded up and
nothing is missing. If any line here is a dealbreaker, close the tab now and keep your $800.
CLIENTSOne. Pepper Partners, an iGaming affiliate network in Tallinn. You would be client #2, and being client #2 of anything is a real risk you are taking on.
ENROLLED46 prospects, one primary contact each.
SENT27 emails. 0 of 27 bounced, 27 of 27 accepted by the receiving server. That is a deliverability number, not a sales number.
REPLIES1. The founder of a fintech company replied "Sure" to an intro offer 27 minutes after the email landed, on 2026-08-31. (Name withheld — he answered a sales email, which is not consent to appear on a pricing page.) The subject line was "a client prospect — you already know the number", and the number was 1.6★ across 240 reviews. 1 of 46 is 2.2%.
MEETINGS0 booked.
CUSTOMERS0 acquired. Ever. Not "0 this month" — this agent has never acquired a customer through outbound in its entire history. Every buyer it has ever had paid before it contacted them. Its own attributable revenue is about $9/mo.
CAPACITY20 sends/day today, limited by domain warmup, the written warm-up table tops out at 50/day. If your model needs thousands of touches, this is the wrong tool and section 06 will tell you so again.
COST TO RUNAbout $2.30/day in model spend. Mentioned because you should know how much of the $800 is machine and how much is a human being paying attention.
So what are you actually buying? Not a track record — there isn't one. You are buying
a mechanic: the audit IS the outreach. The agent computes a real number about each
prospect from public data before it writes anything, and the cold email leads with that number instead
of a pitch. The one reply we have came 27 minutes after a stranger read his own rating in a subject
line. One reply is not a pattern. It is the reason this page exists rather than proof that it works.
// WHY WE MAKE YOU WRITE DOWN WHAT WE MAY NOT SAY
Section 05 asks you for three things most intake forms never ask for: proof points with
a source URL, pre-approved answers to objections, and a never_claim list. That last one is
the field applicants push back on. Here is the honest reason it is mandatory.
An autonomous agent with no never_claim list confabulates. Not maliciously — it fills
gaps, because filling gaps is what generative systems do. This codebase has done it. It once emailed a
real therapist about a website she had never submitted to us. She had not asked for anything. The
sentence was fluent, personalised, and entirely invented, and it went to a real human being with a real
practice. On 2026-08-14 a separate failure sent five cold emails to five real founders with three empty
verdict boxes in them, because a template rendered a field nobody had measured.
Both bugs are now structurally impossible rather than "fixed": copy is templates plus computed slots,
a touch that is missing a field is skipped rather than rendered blank, and the claims
linter fails the whole build if any figure in prose is not sourced. But the linter can only refuse what
you have told it to refuse. never_claim is where you tell it. Empty list, no campaign.
What the linter enforces on every template before a single email can be enrolled, all
of it checked by the same linter that refuses to build a campaign containing an unsourced number:
SLOTSEvery copy slot must resolve to a field your declared signal providers actually emit. Unknown namespace, unknown field, or a proof slot with no matching proof point: build fails.
FIGURESAny number in free prose that is not inside a computed slot must appear in your allowed_literals. If it doesn't, that is a statistic somebody invented, and the build fails. The agent is structurally unable to make up a percentage.
PROVENANCENaming a source platform in prose fails unless the number came from that platform's own API. A rating scraped from a prospect's own page could be anybody's widget; saying otherwise is a fabrication.
FOOTERAn email touch with no postal address and opt-out fails. A non-email touch carrying an email footer also fails, because a mail-merge footer on a LinkedIn note kills the only advantage that channel has.
YOUR WORDSAny touch that speaks for you on compliance, legality, or pricing cannot render until you have supplied that answer verbatim. We do not improvise a legal position on somebody else's behalf.
MODEL USEThe model makes exactly two judgement calls in this entire pipeline: classify an inbound reply, and pick which template a prospect qualifies for. It does not write copy. Ever.
// INTAKE — SIX BLOCKS, EACH ONE A FILE IN YOUR PACK
Required fields are marked *. Nothing here is a sales qualifier
in disguise; every field is read by the code. If you genuinely cannot answer 03 or 05, that is a real NO
and section 07 explains why we would rather tell you now than take a month of your money.
RECEIVED. NOW A HUMAN READS IT.
Your pack draft is in Vlad's inbox. Within 1 business day you get one of three answers:
fit (with a Stripe link and a kickoff date), no fit (with which of
the seven criteria you failed, in writing), or one question — usually about your
trigger signal, because that is the field that decides whether any of this can work.
Nothing is charged before you get that answer, and nothing sends before you approve copy.
Before you submit: this intake is sent through FormSubmit to [email protected] for application review. Do not include passwords, API keys, private customer lists, health data, or other sensitive personal data. Read the privacy policy; contact [email protected] to request access or deletion.
// YOUR PACK, AS THE AGENT WILL READ IT
Built live in your browser from the fields above — nothing is uploaded to render this.
This is the actual shape of clients/<slug>/. If it looks thin, the campaign will be
thin, and it is much cheaper to notice that here.
clients/your-slug/ — draft
# start typing above and this fills itself in
// SEVEN WAYS TO BE A NO
Published because you deserve to know the criteria before you spend 25 minutes, and
because a client we should have refused costs us more than an empty slot does.
NO SIGNALThere is no public, per-company number that predicts need for your thing. If we have to guess who has the problem, the audit has nothing to compute and this is just a sequencer with worse economics than the $99/mo one.
BLOCKED SOURCEThe number exists but the only way to get it violates a robots.txt or a terms of service, and there is no official API. We refuse rather than scrape around it.
NO PROOFYou cannot source three claims. Not "won't" — genuinely cannot point at anything. Then every template is unsourced, the linter fails the build, and there is no campaign to run.
CLAIM IS THE REGULATED PARTMedical outcomes, investment returns, legal results, anything where the sentence that sells is the sentence you need a licence to write. Nothing left to say once we remove what we may not claim.
VOLUMEYou need more than the cap can deliver. 20/day today; the written warm-up table tops out at 50/day. If the math only works at thousands of touches, no.
TRUE PROBLEMYour prospects' trigger signal is describing something real and unfixable about them, and the pitch would be telling a stranger their business is broken. Different product, different company.
CAPACITYYou cannot service even one warm handoff a week, or you want the model to write the copy and the claims to be flexible. Both are hard no. The constraints are the product.
Everything asserted on this page about what the agent has done is
computed from production logs and published daily, including the days that look bad:
/receipts and /this-week.