Product · Artifact Extraction

Nine artifact types. One evidence chain each.

A summary forgets. An artifact doesn’t. Citesvue extracts nine structured artifact types from every recording — each severity-graded, speaker-attributed, and carrying the quote, timestamp, and frame that produced it.

Nine artifact types52:14 UAT · 24 detected
×2
Bug
critical
×3
Requirement
high
×4
Decision
info
×5
Risk
high
×2
Action
medium
×3
Objection
high
×4
Approval
low
×5
Open Q
info
×2
Feature req
low
Taxonomy

The nine types, mapped.

One row per type. What triggers detection, the default severity model, and where the artifact tends to land.

Type
What triggers detection
Default severity
Typical destination
Bug
Observed failure or unexpected behaviour
Critical / High / Medium / Low
Jira · Linear
Requirement
Stated need or constraint
Must / Should / Could
Jira · Notion
Feature request
Suggested addition
Nice-to-have / Strategic
Linear · Notion
Decision
Committed choice with rationale
Binding / Tentative
Confluence · Notion
Action item
Explicit assignment
Owner + due date where stated
Jira · Linear · Slack
Risk
Flagged exposure or concern
Low / Med / High
Confluence · Notion
Objection
Expressed resistance (sales)
Handled / Open
CRM via webhook
Approval
Explicit sign-off
Formal / Informal
Confluence
Open question
Unresolved point
Blocking / Non-blocking
Notion · Slack
Anatomy

Five required fields, always populated. Two optional enrichments.

Quote

Verbatim words that triggered the artifact.

Speaker

Attributed segment-by-segment.

Timestamp

Click-to-jump in the recording.

Frame

Screen state at that moment.

Classification

Type, severity, structured sub-fields.

Two optional enrichments: suggested reproduction steps for bugs and suggested owner for action items — both labelled as suggestions, never asserted as fact.

Pipeline

How extraction actually works.

Transcript segments are scored against artifact-type signatures. Candidates are cross-checked against the visual layer — a bug candidate without an on-screen UI signal is downgraded; an approval candidate corroborated by screen context is promoted.

Severity is inferred from lexical and paralinguistic cues, then surfaced in the review queue for human override. The model is opinionated; the workflow is not.

The review queue

Extraction is a draft, not a verdict.

Every artifact lands in a review queue where a teammate can approve, reject, edit, merge duplicates, or reassign type. Reviewed artifacts build a per-workspace signal that sharpens extraction over time. Nothing ships externally until reviewed — unless the workspace explicitly opts into auto-push.

Worked example

A 52-minute UAT, end-to-end.

Input

1 recording · 52 min · 3 participants

Output

4 bugs · 6 requirements · 3 decisions · 2 risks · 9 action items

Triage

PM reviews queue, adjusts severity on 1 bug, assigns owners to 5 actions, pushes to Jira + Notion in under 4 minutes


The critical bug: a quote from the client at 31:08 about the export producing an empty CSV, paired with the frame showing the error toast, auto-classified critical based on the phrase “this is a blocker for us.”

Structure beats summary

Three lines that explain the difference.

  • A summary compresses.

    An artifact preserves.

  • A summary is one paragraph.

    Artifacts are nine structured tables.

  • A summary dies in a doc.

    Artifacts route into the systems where work happens.

Severity

Severity is a real object, not a vibe.

Every severity assignment exposes the reasoning signal that produced it — the phrase that triggered Critical, the screen context that corroborated it, the confidence score. Reviewers can override; overrides feed workspace-level calibration. Severity is auditable.

Before / after

What artifact extraction actually changes.

Step
Without Citesvue
With Citesvue
UAT documentation time
2–4 hours of manual write-up per session
Under 10 minutes of review
Bug report completeness
Screenshot disconnected from spoken context
Quote + timestamp + frame + reproduction
Decision recovery
“Did we agree to that?” — nobody knows
One queryable decision log per project
On the roadmap

Custom artifact types.

Workspaces will be able to define their own artifact types and extraction rules — for example a compliance obligation, customer commitment, or SLA breach. Tagged as coming, never promised by date.

Common questions

What enterprise buyers ask about extraction.

  • Every extraction surfaces a confidence score. The default workflow assumes human sign-off via the review queue before downstream push — extraction is a draft, not a verdict.
  • Yes, per-workspace. Enable only the types your team actually triages.
  • Yes — CSV, JSON, and full REST API access. Downstream tooling can consume artifacts without going through the UI.
  • Yes, with the caveat that visual grounding (frames, OCR, on-screen evidence) is unavailable for those artifacts.
  • Workspace-level redaction policies are applied before the artifact is written. Audit logs capture every read and export.
Closing argument

Your next recording could be
your most valuable asset.

Or it could sit in a Drive folder nobody opens again. The difference is whether it has citations attached.

  • SetupOne drag-and-drop. No bots, no plugins.
  • First insightCited Q&A on a 60-min recording in under 6 minutes.
  • Cancel anytimeFull data export, full right to erasure.