An AI GTM strategist that can't invent a customer
AI can write a convincing go-to-market strategy in minutes. The problem is what happens a month later, when nobody can tell which parts were known and which were guessed. So I built a GTM system the other way round: every claim carries its evidence, an auditor attacks every draft, the documents must agree with each other before anything moves on, and only a human can lock a decision.
About this project. Built as my own operating system for taking any B2B product to market, whether my own products or a client's. It runs as a set of Claude Code skills and sub-agents, follows a 12-pillar playbook I distilled from a B2B growth masterclass, Alex Hormozi's playbooks and Maja Voje's Go-To-Market Strategist, and can be exported to other agent platforms. The full method draws on licensed material, so it stays private. Its engine (skills, sub-agents, protocol, gates, templates and the Business Profile) is open source and MIT licensed as GTM Machine OS. Each pillar's method is a slot for your own playbook.
Plausible is not the same as true
Ask a model for an ICP, a positioning statement and a pricing page and you'll get all three, fluent and confident. Three things go wrong next. The "customer quotes" in the persona were never said by a customer. The positioning promises something the product can't do yet. And the pricing page contradicts the value proposition two documents earlier, because each was written in isolation.
Strategy built like that fails quietly: the team acts on guesses dressed as findings, and nobody can trace which decision rested on which assumption. I wanted the speed of AI without losing the one thing a GTM strategy is for: knowing what you actually know.
The one-line mandate
Every claim traceable, every document consistent with the others, every decision owned by a human.
Twelve pillars, two gates, one loop
Each pillar is its own skill with a method file, a template and a "done when" list. Each one reads what the earlier pillars decided, and re-running one marks everything after it stale, so nothing downstream rests on an answer that has changed.
Discovery
What the product really does, every public claim checked against it, prices, risks, an exact count of customers and interviews, and the Business Profile that shapes everything after.
Strategy
Segmentation, ICP and buying center, value proposition and positioning, messaging and channels, proof system, packaging and pricing.
Gate 1 — Coherence audit
Seven traces: does every message trace to a value driver, every proof to a message, every price to the category? PASS, PASS-WITH-GAPS or FAIL, with line-level fixes.
Execution
GTM motion first, because it decides how much of the rest applies. Then revenue process and SLAs, demand generation, outbound, enablement and metrics.
Gate 2 — Execution cross-check, then the loop
Do the six plans fit together and fit the hours the team actually has? Then metrics with pre-set thresholds decide when the strategy goes back to pillars 1–3.
Each pillar runs in three roles by default: a researcher that finds external facts and never concludes, a strategist that makes the call and shows the alternatives that lost, and an evidence auditor that attacks the draft from a skeptical-CFO stance. Anything the auditor weakens is fixed or carried into open questions; it never quietly disappears.
Six rules the AI can't talk its way around
They live in one shared protocol every skill loads first, so improving a rule improves every run.
-
01
Every claim is tagged
VALIDATED(customer data or an interview, named),RESEARCHED(an external source, cited with a date) orASSUMED(judgement). An untagged claim is a defect. When most of a document is assumed, its header says it's a hypothesis. -
02
Customer language is never invented
No synthesised quotes, no "a typical buyer might say". If interviews haven't happened, the document says so and the system writes the interview script instead. Once fake quotes are in a file, nobody can tell which ones were real, so this is the one rule with no exceptions.
-
03
The AI proposes, a human locks
Every decision ships as
PROPOSED, with its reasoning and the options that lost. Only the owner marks anythingLOCKED, and a fast run that skipped the audit can't be locked at all. -
04
Numbers, not adjectives
"Improves efficiency" fails review. A number without a source is written
UNKNOWN — needs measurement, and that gap becomes an item in the measurement backlog instead of a hedge in the copy. -
05
Drafts only
The system never sends an email, posts, spends a credit or changes a CRM. It writes the sequence, the spec and the plan, and lists the actions only the owner can take, in order. Paid lookups state their cost and wait for a yes.
-
06
Source gaps are declared, not filled
Where no source covers a method (outreach law per country, marketplace take rates), the skill says
SOURCE-GAPand marks the choice as assumed, rather than inventing a framework and presenting it with the authority of the sourced ones.
A retainer agency is not a SaaS product
The first version gave every business the same advice. Now discovery records a nine-field Business Profile (type, revenue model, deal size, how buyers buy, stage, GTM maturity, markets, capacity and engagement) and no pillar runs until the first five are known. Each pillar applies the defaults the profile sets, and says why when it departs from one.
| If the profile says… | The system defaults to… |
|---|---|
| Before product–market fit | An early customer profile instead of an ICP, a plan to collect proof instead of a proof library, PMF signals before funnel metrics |
| Retainer or project revenue | A value ladder priced between a cost floor and a value ceiling |
| Subscription or usage revenue | A value metric first, then tiers by usage or seats |
| Deal size and how buyers decide | The motion: sales-led for large, complex deals, a marketing- and sales-led hybrid in the middle, product-led for small self-serve ones |
| A one- or two-person team | One funnel, two channels, the smallest tool stack, and outbound volume capped by the hours available |
The failures that became features
Six good documents can still be wrong together
Each pillar can pass its own checks and still contradict the others on prices, names and thresholds, because each was written in isolation. Hence the coherence audit: a fact sweep across every file plus traces from segment to ICP to value driver to message to proof.
A plan is not a result
Building the execution skills exposed a trap: a checklist box like "domains warmed" would tick as soon as a spec described warming them. Now any box that describes something real ("built", "live", "signed") ticks only on the owner's dated confirmation.
The manual and the machine drifted
The human playbook and the skills numbered two pillars differently, and stale notes contradicted the shared protocol. One map now links each pillar to its method, skill and output, and an independent audit fixed 70 inconsistencies in one pass.
Unread sources look like coverage
Three outreach sessions had been extracted but never distilled, so the outbound method looked sourced when it wasn't. The rule now: new material is extracted, distilled into a method file and logged in a coverage table, or it doesn't count.
Built to outlive one tool
The method is plain Markdown. The same system runs in Claude Code with parallel sub-agents, and exports to a Grok Bot as skills, scheduled routines and "ask first" approval rules, so an agent can run the weekly scorecard on its own but still can't send anything without a human.
Where it stands: the strategy pillars have run on two of my products, with decisions proposed and partly locked. The execution pillars are newly built and haven't run on a product yet. The first execution run is next.
Three layers of quality gates
I've argued that evals are the product manager's new quality gate. This system is built around that idea: every output passes cheap mechanical checks first, then a model whose only job is to disagree, then a person.
1 · Deterministic
- A copy linter tests every line against banned phrases and the register of approved claims
- 78 done-when checks, each tick pointing to the section that proves it
- A fact sweep: every price, name and threshold must match across files
- Staleness: change an upstream answer and everything downstream is flagged
2 · Model-graded
- An adversarial auditor returns survives, weakened or refuted for each claim, re-opening the sources
- The coherence audit's seven traces across the strategy
- The execution cross-check: hand-offs, capacity, offer and message fidelity, metric coverage
3 · Human
- Only the owner locks a decision
- Nothing is sent, spent or published by the system
- Real-world steps ("built", "live", "signed") count only on a dated confirmation
Honest status
These are gates, not yet measured evals. Next is a small eval set: test products with planted problems (a fabricated quote, a false public claim, a price that contradicts the value proposition) to measure how often the gates catch them. Until that exists, this page claims no catch rate.
Topics covered
Read next
Store Data Engine
697,299 products from 397 stores in about a week, ~110K enriched with LLMs for ~$31, and +72K more labels propagated free at 97.1% measured precision.
Marketplace Ops · AI · DiscoveryMPB — Warehouse Discovery & 3 Shipped AI Agents
An 11-day product discovery that named an invisible throughput constraint, framed four strategic bets, and shipped three working AI-agent prototypes.
Want a GTM run for your product?
Use the open-source GTM Machine OS with your own playbook, or have me run the full system with you: discovery first, then a strategy where you can see exactly what's known, what's assumed and what to test next.