Skip to main content
When a candidate completes a session, Promptster’s worker pipeline generates several artifact kinds. Each captures a different dimension of the candidate’s work. This guide explains what each artifact contains and how to interpret it.

analytics_v1

Raw behavioral metrics computed from the event stream. Contains 40+ quantitative data points.

Key fields

All rate/ratio fields are between 0.0 and 1.0. Duration fields are in milliseconds.

data_points_v1

Aggregated work-style analysis with behavioral descriptors. Built on top of analytics_v1.

Structure

Process descriptors

The processDescriptors array contains tags that characterize the candidate’s working style:

prompt_analysis_v1

LLM-powered analysis of the candidate’s prompting strategy across five dimensions. Each dimension is rated with a tier and supporting evidence.

Dimensions

Tiers

Each dimension receives one of four tiers:

Example

tool_proficiency_v1

Analysis of how effectively the candidate uses AI coding tools, measured across seven proficiency dimensions.

Dimensions

Scoring

Each dimension receives a numeric score from 1-5 with descriptive reasoning:

key_moments_v1

Identified turning points in the session — moments where the candidate’s approach shifted, a breakthrough occurred, or an important decision was made.

Structure

Moment kinds

test_results_v1

Results from the automated test suite run during submission.
For assessments using the OSS issue library, test results directly indicate whether the candidate fixed the bug. All tests passing means the fix is correct.

candidate_summary

A concise, human-readable summary of the candidate’s session generated by Claude Sonnet.

interview_followups_v1

AI-generated follow-up questions tailored to the candidate’s session, suitable for a technical debrief interview.

Fields

Checking artifact availability

Not all artifacts may be present for every session. Check the kind field to determine what is available:
Sessions with fewer than 3 prompts and 0 file diffs will produce limited artifacts. The analytics_v1 artifact may still be generated, but LLM-powered artifacts like prompt_analysis_v1 and candidate_summary require sufficient data.