A trial project asks a developer to “improve our AI” and ends with disagreement about whether they succeeded. The task never established a baseline, boundary, or reviewable result.
A small paid engineering trial should test how someone delivers a defined outcome, not how much unscoped work they can absorb.
Choose a task with a visible finish
Use one workflow or diagnosis that can be reviewed independently. Examples include reproducing a known integration fault, implementing one isolated operation, or documenting an inherited deployment path.
| Trial element | What to agree |
|---|---|
| Starting state | Repository, fixtures, and known behaviour |
| Required result | One concrete outcome or diagnostic report |
| Exclusions | What the trial will not attempt |
| Access | Minimum approved environment and accounts |
| Review | Evidence and acceptance cases |
Keep payment, rights, confidentiality, and any applicable contractual terms explicit with the appropriate advisers. This is a delivery-design checklist, not a substitute for that review.
Use fictional or properly authorised data
A trial should not require unrestricted customer recordings or production credentials by default. A controlled reproduction can often test reasoning without exposing unrelated business information.
If the task genuinely needs live-system access, define who approves it, what changes are allowed, and how the team protects existing operations. Do not let an interview exercise become an uncontrolled incident response.
Assess the process as well as the final output
Look at the questions the engineer asks, how they separate evidence from hypotheses, and whether they communicate missing dependencies early.
For a fictional duplicate-booking repair, a useful result may include a reproduction, narrowed cause, targeted change, and regression cases. A larger code diff is not automatically a better result.
The debugging-assessment guide gives a scenario you can adapt.
Finish with a maintainable handover
Ask for setup notes, test instructions, known limitations, and a clear account of what was actually verified. Keep unfinished ideas separate from delivered work.
Review the trial against the agreed scope before deciding on an extension. A small success supports a next milestone; it does not prove the person can own every future responsibility without further planning.
The project-brief guide helps turn that next milestone into a complete workflow rather than another vague request.
Paid-trial questions
Should the trial be a complete free product build?
Use a bounded paid task with explicit terms and a reasonable scope.
Must the result include code?
A diagnosis or architecture decision can be useful when the agreed evidence is clear.
What should we decide before starting?
Agree on the outcome, fixtures, access, and review process during the scoping discussion.