For AI labs
"How fast can you staff it?" One week. Here's how
Staffing a task programme in a week is not about hiring fast. It is about filtering fast. A scored trial task on the client's own rubric does most of the work before anyone is onboarded. The schedule, day by day.
Prashant Kumar Singh · · 4 min read
In short
- Day 1: a scored trial task on the client's rubric filters most applicants. Day 2: a one-page guideline and one recorded live session. Days 3–5: a small first batch with every author paired with a reviewer.
- Every day from day 3: pass rate per author, and only those who clear the bar continue.
- The speed comes from the filter, not the hiring. A trial on the real rubric removes most of the onboarding that would otherwise be wasted.
The second question labs ask, after "can you do this", is "how fast". The honest answer for a new domain is one week to a working team producing accepted tasks, and the reason it can be a week is that most of the time in a normal ramp-up is spent onboarding people who will not make it. Filter first, onboard the survivors, and the week is enough. Here is the schedule.
Day 1: the scored trial
Before anyone is briefed, every candidate completes a trial task built to the client's spec and scored against the client's rubric by tests. Not a generic quiz. The trial takes a few hours and most candidates do not pass, which is the point: it removes the people who would have consumed onboarding time and produced rejected tasks. The candidates come from a pool that already exists; day one is about the gate, not the search.
Day 2: the guideline and the session
The people who passed get two things. A one-page guideline: the spec, the end state, what is out of scope, one worked example that passed review, the pre-submission checklist. And one live session, an hour, recorded, where the guideline is walked through and the first questions are answered. One page and one hour, because longer guidelines are not read and longer sessions are not remembered.
Days 3 to 5: the first small batch
Each author writes a small number of tasks, and each author is paired with a reviewer who reads every item and sends written feedback the same day. The batch is deliberately small so feedback arrives before habits form. Questions from the batch go into the guideline that evening, so day four's tasks are written against a better page than day three's.
Every day: the pass rate per author
From day three, every author has a number: the share of their items that passed second review. Authors who clear the bar continue. Authors who do not get one round of targeted feedback and one more small batch. If the number does not move, they leave the programme, with thanks, before their work reaches the client. The team that exists on day six is smaller than the one that started on day one and produces work that is accepted.
| Day | What happens | Output |
|---|---|---|
| 1 | Scored trial on the client's rubric | The shortlist |
| 2 | One-page guideline, one recorded hour | A briefed team |
| 3–5 | Small batch, every item reviewed, same-day feedback | First accepted tasks; guideline v2 |
| Daily | Pass rate per author | A team that clears the bar |
| 6+ | Volume, same review gates | Weekly report |
What the client does during the week
Two things. Approve the trial task on day zero, or send the one they consider a fair test. And answer the guideline questions that come out of days three to five within a day, because those answers become the rubric the volume runs on. Beyond that, the client sees the first batch and the first report at the end of the week, and decides.
Questions people ask
- How long does it take to staff an AI training data programme?
- One week to a team producing accepted tasks: a scored trial on day one, guideline and briefing on day two, a reviewed first batch on days three to five, and daily pass rates deciding who continues.
- How do you vet task authors quickly?
- With a scored trial task built to the client's spec and graded by tests against the client's rubric, before any onboarding. Most candidates do not pass, which removes the wasted onboarding time.
- What does a task-author guideline contain?
- One page: the spec, the end state, what is out of scope, one worked example that passed review, and the pre-submission checklist. It is updated every evening from the day's questions.
If you're a founder
Hiring for a product team follows the same shape: a small paid trial on the real work beats a month of interviews, and a one-page brief beats a long onboarding document.
If you run an AI lab or data team
The week starts with the trial task. Send yours, or let the free review of one of your tasks show you what the trial would look like.
Sources