Every GTM engineer resume now says “Clay.” Some of those people have built a 40-column table that pushes verified, deduped, conditionally-routed records into a sequencer on a schedule. Some have watched a colleague do it. A 30-minute interview will not reliably tell you which is which, because the vocabulary is identical — waterfall, Claygent, HTTP enrichment, lookup. A 60-minute timed build will tell you in one sitting, and it costs you a few hundred data credits to find out.
Here is the exercise we run, the rubric we score it against, what it costs per candidate, and — importantly — why it should sit next to a structured interview rather than replace it.
Why a build, not a conversation
Work samples are genuinely among the strongest things you can put in a hiring loop. They are also routinely oversold. The best current meta-analytic estimate of work-sample validity is .33, not the .54 that gets quoted in every hiring deck — Sackett and colleagues knocked .21 off the old figure once they corrected for systematic overcorrection for range restriction.
That number is the whole argument for how you use this exercise. At .33 it is a strong, cheap signal and a weak oracle. Run it as one leg of a loop alongside a structured interview and a reference check, and you are triangulating. Run it as the single gate and you will reject people who build well but slowly, and hire people who memorised one table.
A timed build tells you whether someone has shipped. It does not tell you whether they can own a function.
The second argument is self-selection. In the Walmart assessment data, applicants who struggled early were 60% more likely to drop out before finishing. That is not a flaw in the instrument — it is the instrument working. A candidate who opens a blank Clay table, sees four named deliverables, and quietly withdraws has saved both of you a 45-minute panel. The pass/fail rubric just formalises what the build was already doing.
The exercise
Fixed table. Pre-seeded rows. No open-ended research. The candidate gets a Clay workspace you own, 25 rows of company domains and contact names already in place, a test Smartlead account, and a written brief. Sixty minutes on the clock, async, screen recording optional.
Leg 1 — the enrichment waterfall. Find a work email for each contact using at least two providers in sequence, with a fallback and a validity check. Not one provider. Two, ordered, with the second only firing when the first returns nothing.
Leg 2 — a conditional formula column. Write a formula column that produces a routing value — say, ICP_tier — from at least three inputs (headcount band, industry, and whether a verified email was found). It must handle blanks without erroring and it must return a clean enum, not a sentence.
Leg 3 — one HTTP API enrichment. Call an external endpoint that is not in Clay’s native integration list. We supply a key and the docs. The candidate has to construct the request, map a nested field out of the JSON response, and handle a non-200 gracefully.
Leg 4 — the Smartlead handoff. Push only the rows that pass the tier filter into a named Smartlead campaign, with at least two personalisation variables mapped to the right fields. The filter condition matters as much as the push.
Four legs maps to four competencies: data sourcing, logic, integration, activation. That is the actual job. Candidates who are strong on three and blank on one are telling you something specific — and usually the blank is leg 3.
Why the waterfall leg is non-negotiable
Because single-provider data decays faster than most hiring managers assume. SaaStr’s read on the current market is blunt: half your CRM is wrong in a world of 49% executive turnover. A GTM engineer who only knows how to call one enrichment provider is building a pipeline on top of data that is already materially stale, and they will not notice until coverage reports come back thin six weeks in. Waterfall fluency is not a nice-to-have on this req — it is the baseline.
What it costs to run
This is the part that surprises people who have been quoted five figures for an assessment platform.
A full enrichment waterfall on a single record costs between 6 and 20 data credits, depending on how many providers fire and which ones. The export leg is cheaper still: pushing a record to an external sequencer is billed as 1 action per record exported. So the arithmetic on a 25-row table looks like this.
| Leg | Unit cost | 25 rows | Notes |
|---|---|---|---|
| Enrichment waterfall | 6–20 credits/record | 150–500 credits | Upper bound assumes every provider in the stack fires |
| Conditional formula column | 0 | 0 | Formula columns are not credit-billed |
| HTTP API enrichment | Varies by endpoint | Your key, your cost | Use a free-tier or internal endpoint |
| Smartlead handoff | 1 action/record | ≤25 actions | Filtered, so usually fewer than 25 |
| Per candidate | — | ~150–500 credits | Trivial against a GTM engineer req |
Even at the top of the range, a five-candidate shortlist costs you roughly 2,500 credits. Compare that to the loaded cost of five 60-minute panel interviews across three interviewers, or to the cost of a mis-hire on a role whose output is the entire outbound motion. Benchmark the comp on the req against our salary benchmarks and the credit spend rounds to zero.
One operational note: give every candidate their own duplicate of the table and cap the workspace. You do not want a candidate looping a waterfall on 2,000 rows because they misread the brief.
The pass/fail rubric
Score each leg binary. No 1–5 scales — they drift, they invite debate, and they let a charming candidate get a 3 for something that does not work. Either the column returns the right value or it does not.
| Leg | Pass condition | Common fail |
|---|---|---|
| Waterfall | ≥2 providers in sequence, fallback fires only on null, validity check present | Single provider, or two providers both firing on every row |
| Formula column | ≥3 inputs, handles blanks, returns a clean enum | Returns a sentence, or errors on any empty cell |
| HTTP enrichment | Request builds, nested field mapped correctly, non-200 handled | Field mapped from the wrong nesting level; no error branch |
| Smartlead handoff | Correct filter, campaign named, ≥2 variables mapped to correct fields | Pushes all 25 rows; variables mapped to display name instead of first name |
The bar: three of four legs pass, and the waterfall must be one of them. Four of four is a strong hire signal. Two or fewer is a no, regardless of how good the explanation is. Publish that bar to the candidate in the brief — it reduces the “was that what you wanted?” email thread to nothing, and it is the fair thing to do.
Have one person score, then have a second person spot-check the two borderline submissions. Consistency across graders is where most work-sample programmes fall apart.
Scoping it so candidates actually finish
Gartner found that most candidates prefer assessments taking between 10 and 30 minutes. Sixty minutes is already double the comfortable ceiling, which means the framing has to work hard.
Three things make the difference:
- Pre-seed everything. Rows, domains, contact names, API key, Smartlead campaign shell. The candidate should spend zero minutes on setup. Open-ended “go find 25 accounts in our ICP” turns a 60-minute exercise into a three-hour one and your completion rate collapses.
- State the deliverables as a checklist. Four named legs, four pass conditions, in the brief. Ambiguity is what makes timed exercises feel hostile.
- Say what happens next. Candidates finish exercises when they believe the exercise leads somewhere. “Pass means a 45-minute structured interview with the hiring manager this week” is worth more than any amount of reassurance about culture.
If you are running this at the top of a funnel rather than mid-loop, cut it to the waterfall and the formula column, call it 25 minutes, and add the API and Smartlead legs to the later round. You will lose some resolution and keep far more of the pipeline.
Where it sits in the loop
Our recommended sequence for a GTM engineer req: a 20-minute screen, then the Clay build, then a structured interview that walks through the candidate’s own submission line by line. That last step is where the .33 gets reinforced — asking “why did you order the providers that way?” surfaces judgment the artefact cannot, and it catches the rare candidate who had help.
For the full competency map, interview questions and comp context, the GTM engineer hiring guide goes deeper than a rubric can. If the role leans further toward infrastructure than activation, the forward-deployed engineer hiring guide is the better reference. And if you are hiring the person who will own the sequencer rather than build the data layer, look at the scoping in marketing operations recruiting instead — different job, different exercise.
The point of all of this is speed. A pre-built table, a published rubric and a binary bar means you can move a shortlist from submission to decision in days, not weeks — and that is usually what decides whether top-tier talent is still available when you make the offer.
