Classic outbound looks for companies that resemble your customers and tells them about a service. This runs the other way: find the companies that already described their own problem in public, and open with their words.
The richest source of those words is job postings. A company that writes "our application is a majestic monolith" paid to publish that sentence and will not walk it back. That is not a guess about pain. It is pain confirmed by a hiring budget.
Everything downstream follows from one split: the event tells you when to write, the data point tells you what to say.
Locating the caller
Ask at most three questions, and stop as soon as the answer is determined.
- What artifact do you already hold? Nothing / an ICP / a list of companies / a scored list / a sequence / a live campaign / replies coming in.
- Has anything been sent from these domains before? Never / warmed but not sent / sending now.
- What is the actual complaint? Only when something is running: no opens, no replies, replies that go nowhere, or good replies that die later.
Do not ask for a strategy deck or a budget. If they had those they would not be asking where to start.
The gates, in one place
Every step below depends on one condition. Route past a failed gate and the work after it is wasted, so name the gate out loud and say what skipping it costs.
| Gate | Condition | If it fails |
|---|---|---|
| G1 — infrastructure | domains authenticated, mailboxes warmed, send caps set | deliverability-audit. With broken sending you cannot tell weak copy from a blocked domain, so every later measurement is unreadable. |
| G2 — a hypothesis exists | you can say who, on what trigger, with what offer | hypothesis-builder. A campaign without a stated hypothesis cannot be judged later, only argued about. |
| G3 — volume | 300 accounts per event-signal hypothesis, 500 per data-point one | bundle related signals, or widen the category. Report the shortfall rather than running a test that cannot produce a readable result. |
| G4 — the list is really the ICP | 90% spot-check passes | back to account-sourcing. Two thirds of raw results are agencies, government and universities. |
| G5 — a reason to write now | rows carry a dated signal or a data point with a source | signal-research. Without it, the no-signal queue has nothing to open with and the campaign is a mass mailing wearing a first name. |
| G6 — contactable | verified addresses, or the LinkedIn route confirmed | waterfall-enrichment. Enrich only what already passed — it is the expensive step. |
| G7 — no meeting ask | every touch ends on an interest question | back to sequence-writer. |
| G8 — the file is clean | final CSV passes the seven-layer check: no suppressed contacts, no invalid addresses, every variable rendered | pre-launch-data-check. One current client or one broken {{first_line}} in the upload costs more than a day's delay. |
The path
Ten steps. Each ends in an artifact, and four of them end in a gate that stops the run.
| # | Step | Route to | Ends with |
|---|---|---|---|
| 1 | ICP and tiers | gtm-market-icp-persona ⧉ |
tier matrix + personas + anti-ICP |
| 2 | Offer and entry rung | offer-factory ⧉ |
5–6 scored bets, entry rung chosen |
| 3 | Signal catalog | signal-catalog |
50 triggers, scored, bundled |
| 4 | Hypotheses | hypothesis-builder |
10 scored, three picked · GATE: volume |
| 5 | Account list | account-sourcing, then signal-research over it |
scored CSV with dated evidence · GATE: 90% ICP spot-check |
| 6 | People and email | waterfall-enrichment |
contacts with verified addresses |
| 7 | Sequence | sequence-writer |
8 touches, both personas · GATE: no meeting ask |
| 8 | Personalization | personalization-pipeline |
two generated fields + confidence gate |
| 9 | Base check | pre-launch-data-check |
GO verdict + rows-to-fix file · GATE: clean file |
| 10 | Launch | deliverability-audit |
SPF/DKIM/DMARC, warm-up, send caps |
Running later: weekly-outreach-report for the weekly read, reply-audit when replies accumulate, ab-test-analyzer for angle tests.
⧉ = lives in a different plugin. gtm-market-icp-persona is in gtm-strategy-skills, offer-factory in gtm-skills. Without them installed, run steps 1–2 by hand — the path still works, it just loses the scaffolding for those two artifacts.
The four rules that carry the whole thing
A state signal does not work alone. "Your framework is out of support" earns "we know, thanks." Pair it with an event — "you raised in March, that framework lost support in March" — and it becomes a conversation. Step 3 forces this distinction; do not let it collapse.
Volume gate: 300 accounts per event-signal hypothesis, 500 per data-point hypothesis. Under that, the test cannot produce a readable result. Bundle related signals under one email theme, or park it. If no bundle clears the gate, the niche was drawn too narrowly — turn the vertical from a targeting filter into a proof selector and widen.
No meeting ask anywhere. Not in email one, not in the breakup, not on LinkedIn. The only job of a cold touch is a reply.
Two generated fields, never a generated letter. A fully generated email drifts to the mean: correct, smooth, faceless. A fixed skeleton holds the voice while the model does the one thing a template cannot — read someone else's document and form a thought about it.
Where runs actually fail
The vertical was used as a filter when it should have been the proof. A niche with strong case studies but only a few hundred companies is an ABM target, not a campaign. Detect the signal across the wider category and show the case study that matches the reader's vertical. Catch this at step 4, not after the list comes back empty.
Two thirds of raw sourcing results are noise — agencies, government contractors, universities. No vendor flag catches them. Budget time for an explicit exclusion pass at step 5, and diff what got dropped: over-filtering quietly removes real accounts.
The quote was never read by a person. Roughly one keyword match in seven is a false positive — "building a legendary legacy" is not a modernization signal. Ten seconds per account before enrichment.
Personalization was attempted inside the sequencer. It cannot be done there. If step 7 produced a first line that carries a thought, step 8 is mandatory, not optional.
The list was enriched before it was filtered. Enrichment is the expensive step. Spend it only on accounts that already passed.
When a live campaign is underperforming, route by symptom
The complaint points at a stage. Do not restart from step 1.
| Symptom | Most likely stage | Route |
|---|---|---|
| Almost nothing opens | infrastructure, not copy | deliverability-audit |
| Opens fine, no replies | the opener has no reason to exist | signal-research, then rewrite the no-signal queue on data points |
| Replies, but all "not now" | targeting is right, timing is not | reply-audit, then revisit signals at step 3 |
| Replies, but all "tell me more" that die | the copy is vague — they are curious about you, not about their own problem | reply-objection-handler (read the tier mix), then the sequence |
| Good replies that go nowhere later | the handoff after the reply | lead-scoring, then reply-objection-handler |
| Numbers look fine, nothing closes | the offer, not the outbound | say so plainly and stop routing — this bundle cannot fix an offer |
That last row matters. A router that always finds an outbound answer is useless, because sometimes the honest answer is that outbound is not the broken part.
Reading the results
| State | Condition | Action |
|---|---|---|
| Kill | replies under 1.5% or interest under 17% | change the signal or the angle, not the copy |
| Optimize | interest 17–30% | new angles, tighter filters |
| Scale | interest above 30% | add a channel, raise volume |
Measure per hypothesis, never per campaign. A campaign mixing two signals reports an average that describes nothing.
Working ceiling on a signal campaign is around 10% positive replies. Vendor-published figures of 35–40% are marketing.
Cost, so nobody has to guess
Roughly 2.5 credits per send-ready lead plus pennies per posting scanned. A 300-lead batch lands near 750 credits and a few dollars of scraping. Say this early — teams routinely assume an order of magnitude more and scope down for no reason.
Enterprise intent platforms starting at $30k/year do not pay back at this volume. Name them, price them, move on.
What never enters the machine
The top 20–30 accounts by score. A person writes those. Automation takes the middle, not the peak.
When someone arrives mid-path
Ask what artifact they already hold, then enter at the next step. Do not restart at step 1 to be thorough — check the existing artifact against its own gate and move forward. A list with no proof quotes goes back to step 5; a sequence with a calendar link goes back to step 7.
Hand off properly
End every run with three lines, no more:
Where you are: [step]
Gate: [met / failed — which one, and what skipping it costs]
Next: run `[skill]` — [one line on what it produces]
Then stop. Do not pre-empt the skill's own questions or summarise what it will say. When someone returns with the same complaint twice, do not route them to the same skill twice — name the assumption that is probably wrong instead.