How we assess GTM engineers.
Candidates build working GTM systems for a fictional company, Northwind Analytics, in sandboxes we run. We test what they built automatically, then they defend it to a founder. Every claim on this site and on gtmuni.com rests on this page.
The scenarios.
About four hours of take-home work in a 72-hour window, one more hour when a role needs a stack scenario, and a 75-minute live session.
| Scenario | Who | Time | What they build |
|---|---|---|---|
| 1. Inbound lead engine Coming soon | Everyone | About 2.25 hours | Webhook → enrich (Vendor A, falling back to Vendor B's bulk API) → dedupe and upsert to the CRM → score → route by territory within five minutes → Slack alert. Plus three SQL questions on the Northwind warehouse. |
| 2. AI research and outbound agent Coming soon | Everyone | About 1.75 hours | An agent researches accounts from Northwind's sources and intent data → ICP score with citations → CRM write-back → a first line → enrols the right contact in a sequence. It must include an evaluation run, a fallback and a human-review queue. An MCP server is a bonus. |
| 3. Your stack Coming soon | When the JD needs it | About 1 hour | One variant: Clay outbound (default), Salesforce Flows and warehouse sync, HubSpot workflows, data and dbt, product signals to PQL, or call intelligence. An optional CRM clean-up extension for data-hygiene roles. |
| Live session | Everyone | 75 minutes | System design (15) · an incident drill on their own build (15) · a defence, including "when is Clay enough and when do you write code?" (20) · a demo to Northwind's VP Sales (10) · debugging a predecessor's script (15). |
The environment.
Some tools are real and some are our simulations. The simulations follow each vendor's API rules: auth and token expiry, required fields and real error messages, duplicate rules, pagination and rate limits. We say which is which, here and in every report.
| Piece | Type | What it does |
|---|---|---|
| CRM | Simulated | Salesforce-compatible or HubSpot-compatible, matched to the JD and seeded with Northwind data |
| Contact and company data | Simulated | Apollo-style and ZoomInfo-style APIs: API keys, OAuth2 with token refresh, pagination, rate limits, bulk jobs |
| Sales engagement | Simulated | Outreach-style sequencer: enrol and pause, open and reply webhooks |
| Signals | Simulated | Intent data, call transcripts and product usage events |
| Research sources | Simulated | About 10 pages per account, with planted facts and traps |
| Build tools | Real | n8n, a GitHub repo (Python or TypeScript), Clay or Make |
| Warehouse | Real | Postgres, one schema per candidate |
| LLM | Real | Claude or OpenAI through our gateway, with a spending cap and every call logged |
| Slack | Real | One channel per candidate |
The four kinds of simulated API cover all six ways GTM tools connect: OAuth2 with refresh, bearer token, API key, basic auth, webhooks and bulk or async jobs. Each ships with vendor-style API docs.
How they submit.
A live endpoint
Something our tests can trigger: their n8n workflow, deployed code, or a Clay or Make webhook kept live for 72 hours.
The build
An n8n export, the final repo commit, or a shared Clay table or Make blueprint.
The explanation
A README, a five-minute recorded walkthrough and a short design note, including how they used AI tools.
How it's scored.
Automated tests
Our test runner fires planted data at the build (duplicates, bad fields, rate limits, expiring tokens, trap accounts), checks the CRM, API logs, Slack and LLM logs against an answer key, then replays everything to confirm nothing duplicates. The agent is tested on 10 hidden accounts.
Eight skills
Every check is tagged to a skill: Integration; Enrichment & data quality; CRM systems; AI agents; Data & SQL; Code; Reliability & diagnosis; Commercial judgment & communication. Each is scored 0–100 in bands: Expert, Proficient, Competent, Developing, Not yet demonstrated.
Code level
No-code builder, Scripter or Engineer, shown separately.
Cannot-fail checks
No credentials in code, no one contacted twice, no do-not-contact person enrolled, no silently lost records, and the candidate can explain their own build.
Assessment brief
A founder reviews every build, runs the live session and writes the brief at the top of the report.
Built from the job data.
The scenarios cover every skill that appears in more than about 45% of US GTM engineer job descriptions. We built them from more than 200 US GTM engineer job descriptions analysed in 2026.
| Skill or tool | Share of JDs | Tested in |
|---|---|---|
| AI agents | 72–73% | Scenario 2 |
| APIs and webhooks | 64–76% | Scenarios 1 and 2 |
| Outbound and sequencing | 68–69% | Scenario 2, Clay outbound variant |
| Enrichment and data hygiene | 64–68% | Scenario 1, clean-up extension |
| Code required | 59–68% | Code track, debug task |
| LLMs (Claude, OpenAI) | 61–64% | Scenario 2 |
| SQL | 46–61% | Scenario 1 |
| Clay | 51–54% | Clay track, Clay outbound variant |
| Python / TypeScript or JavaScript | 38–54% / 26–49% | Code track |
| HubSpot / Salesforce | 38–44% / 18–59% | CRM matched to the JD |
| Lead routing / speed-to-lead | 26–49% / 13–37% | Scenario 1 |
| Coding agents | 24–33% | Allowed; explained in the design note |
| Customer-facing work | 24–31% | Stakeholder demo |
| Data warehouse | 18–29% | Data variant |
| Gong | 5–20% | Call-intelligence variant |
| MCP | 5–10% | Scenario 2 bonus |
Source: 80 live JDs read in full and checked against company job boards (November 2026, companies up to 200 employees), plus about 142 requisitions analysed in August 2026. Ranges compare roles open 60+ days with recent ones.
What this does not cover.
Simulated tools are not the vendor's own product, so a JD-match page marks a tool "Demonstrated", "Discussed" or "Not assessed", never "tested live" when we simulated it. CPQ has no free sandbox and is covered only in the live session. The report's last page lists the limits for each candidate.