GTM Expert

Onboarding — start here

What this does. You bring a go-to-market question. This skill decides which of six practitioners actually has a tested answer, reads their distilled reference, and returns the method with its real numbers — plus an honest note on how well-evidenced it is and where another of the six would tell you the opposite.

The six corpora were researched independently and each carries an honesty layer: which claims are measured, which are asserted, which are course marketing. That layer is the reason this skill exists. GTM advice is abundant and mostly unfalsifiable; what is scarce is knowing which parts of it someone actually ran.

What it does not do, and where to go instead.

• It does not research a specific company. It knows methods, not accounts. For who a company sells to, use icp-research. For their live ad targeting, use ad-library-recon. For firmographics, use Blitz or the Clay MCP.
• It does not find or enrich leads. It will tell you how Crawford builds a pain-based list; it will not build one. Use blitz-gtm-brainstorm → blitz-create-script, or clay-api, or lead-generation.
• It does not write the copy, the post, or the sequence. It gives you the structural constraints — step count, cadence, what the opener must not do — and then you write. For LinkedIn content production specifically, the linkedin-content-creator and uppr-linkedin-content skills own that.
• It does not know Grow Steady's own ICP. That lives in growsteady.md at the repo root and is a separate artifact. Read it yourself when the question is about this business rather than about GTM in general.
• It is not current. The corpora were compiled August 2026 and freeze there. For anything algorithm-dependent (LinkedIn reach, deliverability rules) treat the numbers as a starting hypothesis, not a live reading.

✅ Prerequisites — nothing to install

No API keys. No MCP servers. No network calls. Nothing to sign up for. Unlike icp-research (which requires both Exa and Firecrawl connected), this skill is entirely self-contained: six local markdown files under references/. If someone tells you this skill needs Exa, Firecrawl, Clay or Blitz, they are confusing it with a different skill — it reads no live data at all, by design.

The one dependency is that the six reference files exist and are populated:

wc -l .agents/skills/gtm-expert/references/*.md

You should see six files, each roughly 195–280 lines. If any is missing or near-empty, the distillation did not complete — regenerate it from the full corpora in .agents/_research/gtm-experts/.

Verify. Open references/jordan-crawford.md and check the "Numbers and thresholds" table has rows in it. That table is the densest part of the skill; if it is populated, the rest almost certainly is.

How to use it

Just ask a GTM question in plain language. You do not need to know which of the six experts covers it — working that out is the skill's whole job:

• "How should I structure a cold email sequence for a €25k ACV agency offer?"
• "Our reply rate dropped from 4% to 1.5%, what do we fix first?"
• "Is 3 posts a week right, or should I be posting daily?"
• "Should we hire an SDR or deploy an AI SDR?"
• "Sanity-check this GTM plan" (paste the plan)
• "What's a defensible ICP for a 12-person data agency?"

To force it explicitly: /gtm-expert <your question>

What you get back: the method with its real steps, the numbers with their units and denominators, an explicit label on how well-evidenced each claim is (measured / asserted / marketing), and — where one exists — the disagreement between experts plus the axis that resolves it. If the question falls outside what the six cover, you get told that instead of a confident guess.

Two things worth knowing before you rely on it. The corpora freeze at August 2026, so anything algorithm-dependent (LinkedIn reach, email deliverability) is a starting hypothesis rather than a live reading. And every one of the six sells something downstream of their advice — the skill discounts for that explicitly, and you should too.

Cost. Free to run, but not weightless in context: each reference is 195–280 lines. Reading one is cheap; reading all six is not. The router below exists so you read one or two, not six.

Rest of the skill. Six references, one per expert, each with a Consult for / Do not consult for header so a wrong route is caught in one line. The full uncompressed corpora stay in .agents/_research/gtm-experts/ — go there only when a distilled reference points you to a section it had to cut.

Route the question

Match on what the question is about, not on vocabulary. Most GTM questions arrive dressed as something else.

When two experts are listed, read the first one's section, then check the second only for contradiction. Reading both in full is usually waste.

When nothing matches, say so rather than stretching. These six cover outbound, content/demand, 0-to-1 GTM, sales ops, and LinkedIn. They do not cover paid search, SEO technical work, PLG/product analytics, channel/reseller economics, or anything B2C. Answer from general knowledge and label it as such.

What all six agree on

Convergence across six independently-researched practitioners with different businesses and different incentives is the strongest evidence this skill holds. Treat these as safe defaults; the burden of proof is on deviating.

1. Targeting beats copywriting, and it is not close. Lieben claims 3–5x from a signal-defined list over a firmographic one with worse copy. Vacca's Rule 98 calls copy optimisation on a broken list the single most common wasted effort in outbound. Crawford: "finding 5,000 wrong records is more valuable than enriching 50,000 records." If reply rates are bad, the list is the first suspect, not the subject line.

1. Relevance beats personalisation. A first-name token is a mail merge. Vacca reports false personalisation ("your impressive background") performs worse than none; Crawford says unlimited personalisation reads as surveillance. The unit of relevance is a problem the recipient has, stated in terms they would recognise.

1. Firmographic ICPs are not ICPs. Crawford: "'50 to 1,000 employees in vertical X' — that's not an ICP, that's an Apollo filter." Voje calls firmographics "very irrelevant" to early segment choice. Both build from problem, not from size.

1. Own the workflow, rent nothing load-bearing. Vacca: "a workflow you own is worth more than a subscription you rent." Stam: buy for API/CLI/MCP portability, closed vendor UIs are rented. Crawford built his own agent rather than deepen a platform dependency. Scoring logic in a markdown file is auditable; scoring logic in a vendor UI is not.

1. AI amplifies a system; it does not create one. Vacca's Rule 100 and Rule 95 ("nothing fixes a broken process by putting a person in front of it — applies to AI SDRs too"), Stam's "AI is a multiplier, not a corrector," and Hérubel's attack on 100-pieces-in-a-day content all say the same thing from three directions. Bolted onto an unowned process, AI scales the mess at machine speed.

1. Activity metrics are lying to you. Hérubel replaces MQLs with forecasted closed-won revenue. Vacca: 50 meetings and zero closed deals is a targeting failure, not a win — track cost per opportunity. Stam rejects meeting-count as deal progress and the average sales cycle as a forecast yardstick (use the median win cycle; losses drag the average long).

1. Smaller and more frequent beats bigger and rarer. Vacca: sub-50-recipient campaigns reply at 5.8% against 2.1% for 1,000+ blasts; frequency beats reach (8x to 1,000 accounts over 1x to 8,000). Lieben: micro-lists of 500–1,000, campaigns under 5,000 hyper-enriched prospects. Stam: "a list is not pipeline."

Where they genuinely disagree

These are real conflicts, not phrasing differences. When a question lands on one, surface the disagreement and the axis that resolves it rather than picking a side silently — the user's situation is usually what decides it.

Signals: edge or commodity? Voje calls signal-driven outbound the current edge. Vacca says signal density outranks company size and that three likes on your own post beat any third-party intent signal. Crawford refuses signals-based outbound outright, and his reason is structural: "No great product gets worse the more people use it. Signals do." He says GTM alpha from a shared signal halves year over year. Resolving axis: is the signal proprietary or purchased? Crawford's objection is to bought, shared signals — hiring data everyone has. Vacca's best example (engagement on your own content) is proprietary and Crawford would not object to it. If the signal is on a vendor's price list, Crawford's decay argument applies.

Rank accounts by wallet, or by pain? Voje reverse-engineers ICP on ARR/LTV — she flags this herself as the position product leaders come after her for. Crawford says ranking by wallet size is wrong: rank by problem size relative to wallet, because "a small company drowning in a problem you solve beats a giant who barely feels it." Resolving axis: are you optimising for the next quarter or for a repeatable motion? Voje writes for bootstrapped companies needing revenue now. Crawford writes for teams building a durable list.

How many follow-ups? Voje: two, not seven — explicitly against the 5–10 touch orthodoxy. Lieben and Vacca both run multi-step sequences (see their references for exact structures). Hérubel attacks the premise: linear sequences run 4–7 weeks against B2B cycles of 4–7 months, and hard CTAs trigger psychological reactance (he cites Steindl et al., 2015). Resolving axis: is there a parallel always-on channel? With content and intent monitoring running, Hérubel's non-linear model is available and step count matters less. Without one, the sequence is all you have.

AI SDRs. Stam is the only one of the six who reports deploying one at scale with a positive result (80% AI-SDR-influenced inbound demos — first-party, unverifiable). Voje surfaces a survey quote of six months producing zero opportunities and is sceptical of "agents" that are n8n workflows. Vacca and Crawford both say the role is being automated but that automating a broken process just breaks it faster. Resolving axis: does a human currently do this job well? Every one of them agrees AI should automate a working process, never substitute for a missing one.

Job postings as a signal. Crawford contradicts himself here, and knows it: job data built his business, and his own playbook copy now names it the canonical bad signal ("I see you're hiring compliance people — everyone sees this"). His reconciliation: weak as a shared opener, strong as a reading signal — what a company cannot do yet, in its own words. Use it to understand, not to open.

Inbound vs outbound. Voje published "inbound is out, outbound is in" — in a post sponsored by ZoomInfo, and her own unsponsored survey a month later found inbound the most common motion overall and no motion correlating with growth rate. Hérubel calls inbound-only "an ego position that halves your ROI." Treat the sponsored headline as the weaker claim and the combine-both position as the consensus.

Discount for commercial interest

Every one of the six sells something, and in each case the thing they sell sits directly downstream of the advice. This is not a reason to discard them — they are practitioners because they sell it — but a claim that happens to sell the author's product needs harder evidence than one that costs them.

Two of them are honest against their own interest in ways worth crediting: Vacca publicly argued AI outreach would kill agencies and then launched one anyway, saying so; Crawford killed his own "16x cheaper than OpenAI" post after an adversarial review found the ranking did not survive.

How to answer

Once you have the reference open, the shape of a good answer is the same every time.

1. Name whose method it is, and say what kind of claim it is. "Lieben's tiering system" is useful; "best practice says" is not. Every reference marks claims as measured, asserted, or promotional — carry that through. An unlabelled number gets treated as fact by whoever reads your answer next.
2. Give the actual steps, not the summary of the steps. A six-step method compressed to a sentence is a slogan. If the reference has the steps, reproduce them.
3. Keep numbers with their units and denominators. "5.8% reply rate on sub-50-recipient campaigns" survives being quoted; "5.8%" does not.
4. Surface a conflict when one exists. If another of the six would tell the user the opposite, say so and give the resolving axis. Suppressing that to sound decisive is the main way this skill could do harm.
5. Say what would falsify it. Most of these methods have a cheap test — a 50-recipient campaign, a single positioning interview, one week of a different posting cadence. Naming the test is more useful than defending the claim.
6. Refuse gracefully when it is out of scope. The corpora freeze at August 2026 and cover a specific slice of B2B. Answering a paid-search or PLG question from these six and implying it is grounded is worse than saying it is not covered.

Two worked routes

"Our cold email reply rate dropped from 4% to 1.5%. What do we fix?" Route: Lieben (infrastructure and deliverability) first, Vacca second (Rule 98). Do not start with copy — consensus point 1 says the list and the infrastructure are the first suspects, and Lieben's position that cold email deliverability is structurally only 60–70% and worsening reframes the question: part of the drop may not be yours to fix. Return his deliverability thresholds, his list-size guidance, and the diagnostic order. Flag that his headline reply benchmarks are the weakest-sourced numbers in his corpus.

"We're a 12-person agency at €25k ACV. Should we build a community?" Route: Voje (motion selection) first, Stam second (community and events as a channel). Voje's specific position — do not build, penetrate existing communities, because building is a cold-start problem — is the direct answer. Flag her "worked well for 97% of 650+ companies" claim as the weakest number in her corpus: no definition of "worked well", no methodology. Then give her beachhead criteria so the question becomes which community.