The Schneider Manual

How the marketing-agent method works, what we built from it, how to run it, and what it costs. Written from a 44-minute masterclass by Cody Schneider, then from actually building the thing and finding out where it bends.

1. The one idea

Target what someone just showed interest in, not what industry they happen to work in.

Cold email is collapsing because AI-generated volume has flooded every inbox. Sending more into that is not a strategy. The way through is to only contact people who have just raised their hand.

On LinkedIn, a like or a comment is that hand raise. It says a person cares about a subject right now. Firmographics tell you who someone is; engagement tells you what they currently want. So: watch the accounts your customers read, collect the people who engage, qualify them, find their contact details, and write about the thing they just engaged with.

2. The rule that governs cost

In our build the model is called in three places, and never once per lead in the pipeline stages:

Measured on 962 real people, stage 1 resolved 333 with no model call at all. That saving compounds every single day the system runs.

3. The pipeline, stage by stage

Stage 1 — Choose the source accounts

10–20 accounts whose audience is your customer. Company pages count as much as people — most teams collect only individuals and halve their coverage. Pick them by hand. Cody deliberately does not automate this, and neither do we: a business already knows who its customers follow.

We seeded ours from the knowledge bases, taking each company's LinkedIn URL from its own website rather than guessing it from the name. Guessing produces a page that looks right and silently tracks the wrong firm.

Screen the accounts before you track them. We did not, and paid for it — see the employer-affinity finding in section 9.

Stage 2 — Collect the engagers

Daily: find each tracked page's new posts, pull everyone who reacted or commented, dedupe on the public profile URL. One Apify actor run per company — the company-posts actor scrapes reactions inline, so this is one run rather than one per post.

Stage 3 — The ICP gate, and it goes BEFORE enrichment

Our gate asks two questions rather than one:

• Role — does this person specify, build, or buy things?
• Domain — are they plausibly in our field?

Both true is a yes. Role but unclear domain is the ambiguous middle, and that is where stage 2 earns its cost. Domain vocabulary lives in each campaign's config, so magnetics and home building bring their own without touching the classifier.

Current state on the full harvest: home building 135 yes / 522 maybe / 385 no out of 1,042; magnetics 72 / 665 / 225 out of 962. The large middle column is the gate refusing to qualify on role alone. Emptying it is what stage 2 is for.

Stage 4 — Waterfall enrichment, cheapest first

Providers in ascending cost order, only the misses falling through: GitLeads → Apollo → Origami or Prospeo, with LeadMagic for mobile numbers. Roughly an 80% find rate is realistic.

Stage 5 — Verify before sending

Every address through Million Verifier. Enrichment output is not clean enough to send to, and bounce rate is the fastest way to destroy a sending reputation.

Stage 6 — Sending infrastructure

Four domain lanes, kept apart: burner domains for cold, separate for email marketing, separate for transactional, and the real business domain which never sends cold. Ten thousand cold sends from a live domain ruins it — and at our scale that damage would land across the whole network.

Stage 7 — Replies

The sending platform fires a webhook on reply. Classify intent, draft toward a booking link, and connect the calendar back so the system sees whether a meeting actually happened rather than assuming.

4. From a list to a department

The campaign layer lives in /opt/autoengine/campaign.py, beside the engine rather than inside it, and runs on first activation and on every daily run. It is entirely deterministic — no model calls at all. That is deliberate: a business owner has to be able to ask why a proposal says what it says and get the same answer tomorrow.

It derives the angles the site can actually evidence

Seven candidate propositions — local, speed and guarantee, proof, expertise, price, range, risk-reversal — each declaring what the site must carry before the claim may be made. A speed claim needs the site to already promise speed. A proof claim needs named customers or review data. An angle with no support is not written. A site with no signals at all yields exactly one angle, not seven invented ones.

It writes real creative, four formats per angle

• Search ad — written inside Google's actual 30-character headline and 90-character description limits, with the character count shown. Copy that ignores the limit is truncated mid-word in the auction, so the generator returns fewer headlines rather than clipped ones.
• Social post — hook, body, call to action, and an instruction to use a real photograph, because a stock image reads as one.
• Video — a 20-second vertical shot list, six beats, shot on a phone. The first three seconds do the work; production value does not.
• Landing page — the page the ad must point at, with three proof points and a note to replace any that cannot be evidenced.

It proposes spending, and asks first

In the shape Tim described: “I want to spend $250 in Google to find out the acquisition cost” → comes back with “it is $32 a lead” → the owner decides. Every proposal names the channel, the amount, the duration, the question it answers, and the rule that stops it. A proposal without a kill condition is a bill.

Both limits are enforced, not advertised

A cost-per-lead cap and a monthly ceiling, exactly as Tim put it: “I will pay $5 a lead, and do not bill me more than $1,000 a month.” An approval that would breach the ceiling is refused with the arithmetic shown. A measured cost per lead above the cap returns a kill verdict rather than a suggestion.

Free channels first — as a gate, not a preference

Dawn Jordan arrived at the same rule from the other direction on her first day of testing: “assume we have five dollars and we are starving — how do we get a customer? Ask for ten thousand and someone will suggest a billboard.” So free is the default posture, and staying there is presented as a legitimate outcome rather than a trial.

Onboarding recommends rather than interrogates

Also Dawn's: asking a non-marketer to choose goals and channels is asking them the question they hired you to answer. POST /recommend reads the live audit and answers it — goals and channels already chosen, each carrying the signal it came from. On a real magnetics site it returns “Direct sales, because the site lists 38 products” and “Be findable at all, because there are only 229 words on it.” Everything stays editable: a recommendation you cannot interrogate is a default wearing confidence.

What it is tested against

Twenty assertions cover the guardrails — free-first refuses, the ceiling refuses, the cap judges, a channel with no leads reports no number, and a site with no signals invents no angle. All pass. The layer was then exercised against 259 real workspaces with zero exceptions, because a generator that only works on the site it was developed against breaks 189 others at 6am.

5. Knowing a lead happened

/opt/wt-leads on port 8956, proxied at /l/ on the client's own domain — so no CORS, no third-party script on a site we manage for someone else, and nothing for an ad blocker to object to.

Two things are recorded, and the difference between them is the point

Attribution is captured on arrival — utm_source, medium, campaign and content, falling back to the referrer host — and held in sessionStorage for the visit. No cookie, no fingerprint, no identity. The IP is hashed server-side purely to rate-limit. What is stored is: something was tapped, at a time, by someone who arrived from somewhere.

Verified end to end in a browser on 13 August: a tap on the phone number arrived attributed to google / cpc / wells-test / local-angle. Test rows were then deleted so the real count starts clean.

6. A knowledge base for any trade

Maps for local trades, LinkedIn for corporate industries

kb_from_maps.py and kb_from_linkedin.py each turn a trade name into a knowledge-base seed. Which one you reach for is not a preference, it is a fact about where that industry exists:

• Google Maps for a local trade. Every business carries a category from Google's own fixed taxonomy, so “Water well drilling service” is a distinct category from “Oil field equipment supplier.” Relevance is decided by Google's classification of the business rather than by pattern-matching its marketing copy — the difference between asking does this text mention wells and what is this business.
• LinkedIn for an industry with genuine corporate presence. Magnetics and home building both work here. Water-well drilling does not, at all.

The vocabulary comes from what the industry actually posts

Maps and LinkedIn both tell you who exists. Neither tells you what the trade talks about. That turned out to be sitting unused in data we already had: every engagement row carries the URL of the post that drew it, and a LinkedIn post URL encodes the author's own hashtags.

topics_from_engagement.py reads them and splits the result three ways:

Magnetics gained fourteen topics this way — Rare Earths weighted 197, Critical Minerals 117, Magnetic Separation, Tramp Metal, Heavy Rare Earths. PolyMagnetics now briefs “The Complete Guide to Rare Earths” instead of something generic.

7. How to run it

# a full dry pass -- touches nothing, proves the wiring
cd /opt/wt-outbound
python3 run.py --domain magnetics.wholereach.com --dry

# harvest engagers across every tracked company page
python3 /root/harvest_all.py        # magnetics
python3 /root/harvest_builder.py    # home building

# qualify by role and domain (breadth ranking is kept, but see section 9)
python3 /opt/wt-outbound/icp_gate.py

# pull the industry's own vocabulary out of the harvest
python3 topics_from_engagement.py magnetics --harvest harvest-magnetics.json
python3 topics_from_engagement.py magnetics --harvest harvest-magnetics.json --write
python3 /opt/autoengine/kb/gather-kb.py magnetics

# build a knowledge base for a trade we have never covered
python3 kb_from_maps.py waterwells --location "Texas" --write

# give already-activated sites the campaign slate (idempotent)
python3 /opt/autoengine/backfill_campaign.py royallswindmill.com --write

# the guardrails -- run these after any change
python3 test_pipeline.py                 # pipeline
python3 /tmp/test_spend.py               # campaign caps, free-first, cost per lead

It runs itself daily at 06:25, before the huddle, via /root/daily-harvest.sh. Yesterday's results are moved aside so new engagement is seen, and fourteen days of snapshots are kept.

8. What is built and what is blocked

Every unbuilt adapter is written and dry — it does nothing without its key and works the moment one exists. Nothing spends money by accident: an adapter goes live only when its key is present and dry-run has been explicitly disabled. Two conditions, not one.

9. What we learned the hard way

An actor returned the wrong company's data, silently

The first LinkedIn actor we tried — the publisher named in the video — was a profile scraper. Handed a company page it did not error; it returned Satya Nadella's posts. Plausible data for the wrong entity.

The engine served one client an entirely wrong industry

A water-well drilling client was shown a competitive landscape of Clearscope, Surfer and Frase — SEO software companies. Nothing was fabricated: every company real, every description accurate, the whole thing about the wrong field. That is harder to catch than invention, because nothing looks wrong.

Two causes, compounding. No water-well knowledge base existed. And the engine had read the site's niche as “Royall business” — the company's own name, not its trade — so no keyword could ever have matched.

Breadth is near-worthless as a signal in a trade

We first ranked people by how many different companies they engaged with, on the logic that following several proves industry membership. It surfaced sales staff. We almost concluded the method does not work for our field.

The sort was wrong, not the method — and on the full harvest the size of the error is now measurable. Of the people who genuinely qualify, 86% in home building and 90% in magnetics engaged with exactly one company.

Most corporate engagement is employees liking their employer

55% of home-builder engagement is employer affinity — #ilovelennar, #thisistmliving, #webuildpeopletoo, national intern day. Those pages are recruiting channels, not trade publications. It also explains why the qualified builder list filled with sales and HR staff rather than buyers.

Magnetics runs the other way: 64% genuine subject matter. Worth measuring before pointing the audience engine at any consumer-facing vertical — and worth screening source accounts on.

Half a local trade has no website at all

Of 56 water-well businesses found in Texas, 31 have no working website — 23 list none at all and 8 list one that does not load. That is 55%, and it is not a gap in the data, it is the state of the trade.

It is also the clearest sales argument the engine has: a business with no site cannot be marketed to by any competitor of ours either, because every marketing tool on the market starts by asking for a URL. The full finding, with the prospect list.

LinkedIn is the wrong instrument for a local trade

Searching it for well drilling returned Tenaris, Schlumberger and 52 other oil-and-gas multinationals. A two-truck outfit in Luling has no company page there at all — the population we wanted is simply not in that corpus, and no filter fixes a missing population.

Statistical significance is a real ceiling, not an engineering gap

A $250 pilot at $32 a lead produces eight leads. You cannot distinguish two creatives from eight leads; telling a 3% conversion rate from a 5% one needs hundreds of conversions. So “test 25 creatives and tell me which works” requires either a much larger budget or a much longer run.

No amount of building fixes this. What a $250 pilot honestly delivers is an approximate cost of acquisition and the ability to kill anything obviously terrible. Promising the first and delivering it beats promising the second.

Read the schema, do not guess the payload

Our first working actor returned nothing because we sent urls where it expected targetUrls. Reading each actor's published input schema also revealed scrapeReactions — which halves the number of runs. Guessing cost an hour; reading saved the budget.

A lexicon must never be built from its own output

The vocabulary extractor splits run-together hashtags back into words using the knowledge base as its dictionary. An earlier run had written raw tags into that same knowledge base as topic names — so rareearths became a dictionary word, matched whole, and never split. It now reads only from prose and ignores anything beginning with a hash.

10. What is easy, and what is hard

Easy — days, and mostly a credential rather than a design problem

Hard — and three of these are where a system like this usually starts lying

1. Knowing a lead happened, properly. Section 5 captures the tap. The call itself needs a tracked number, conversion tracking on a site we may not host, and attribution back to the creative. Several weeks. Every platform that quotes you a cost per lead is making quiet assumptions here.
2. Significance at pilot budgets. See section 9. Not solvable by building.
3. Running and optimising paid campaigns. Placing the ad is bounded work. Managing bids, negative keywords and knowing when to kill a loser is judgment. About a month for an honest first version, and expect it to be mediocre until it has run on real accounts.
4. Publishing to sites we do not host. We write directly to our own network with a snapshot and rollback on every change. Wix, Squarespace or WordPress elsewhere needs a connector each — a week apiece, and they break when the platforms change.
5. Producing the images and video. The engine writes the shot list. It does not hold a camera. For a trade, a phone photograph of real finished work beats anything generated — so this may be something we ask the client for rather than build.
6. Vocabulary at real depth. We now extract what an industry hashtags. We do not extract what it argues about. That needs trade publications, forums, and ideally recordings of real sales calls — the difference between an article that is correct and one worth reading.

11. What it costs

$39 a month is what is actually being spent today, across the whole network. Every deferred tool together would add about $410. The engine itself costs nothing per client to run — no model calls in the routine path — which is the single largest cost control in the design and the reason a $140 tier is possible at all.

The full ledger, with what each tool has to return to earn its place. Key state there is read from the machine at build time rather than from memory.

12. The second engine

The other half of the method turns real conversations into published content. Calls, interviews, transcripts, lost deals — the source material becomes posts.

Never ask a model to “write good content.” With nothing to work from it produces the average of everything it has read, which is precisely what platforms now detect. Every piece must start from something a human actually said.

For this network it runs on topic pages rather than personal brands — accounts named for a subject, not a person. That fits our institutional-byline rule exactly, and the evidence says it works at least as well anyway.

The standing question on the huddle page every morning is this engine's intake: one question a day, rotating, because the knowledge that matters keeps arriving.

13. Where everything lives