Emiliya Ivanova-Nikolova
Task

2026 · In progress

Seven agents that turn business apps into training videos

I'm building a framework of seven AI agents that records a live business application and creates how-to training videos with voice-over, then waits for my approval. It has produced four videos so far.

  • AI agents
  • Playwright
  • Voice-over
  • Instructional design
  • QA

The problem

Internal teams need how-to training videos for the business applications they use. Those videos have to match the real application and the latest documentation, and they have to be revised when either one changes.

The result so far

The framework has produced four videos with voice-over. The videos are confidential, so I can’t show them here.

How the agents work together

The agents communicate with each other, and nothing is final until I approve it. The framework has two skills: create training videos and revise training videos. Creating a video takes 10 steps, with two points where the framework stops for my approval.

  1. Steps 1 and 2scenario-analystReads a scenario file and related resources, then extracts the goal, audience, prerequisites, steps, rules, terminology, and expected outcome. Drafts the video brief and flags any gaps.
  2. Step 3instructional-designerTurns the analyzed scenario into a focused learning design: one objective, scope, scene structure, and duration, plus where highlights, zooms, callouts, and transitions help. Finalizes the brief.
  3. Step 4script-storyboard-writerWrites the on-screen caption script and a detailed storyboard as structured JSON, then exports Markdown versions for review.
  4. Step 5app-interface-specialistChecks the storyboard against the real application and its documentation, including labels, buttons, fields, status values, navigation, and business rules. Uses read-only Playwright inspection, and fixes discrepancies and outdated-video risks.
  5. My approval of the storyboardI review the storyboard. If something needs fixing, the agents revise it and ask me again.
  6. Steps 6 and 7video-producerWorks from the approved storyboard. Runs the Playwright capture of the real application, renders a captioned H.264 MP4 and SRT file, and runs technical validation.
  7. Steps 8 and 9qa-reviewerIndependently checks scripts, storyboards, clarity, sync, readability, and data protection. Issues pass or fail findings with specific corrections, and rechecks after fixes.
  8. My approval of the videoThe framework waits for my decision. If something needs fixing, the agents revise it and ask me again.
  9. Step 10video-producerPackages the approved deliverables.

Revising an existing video

Revisions start with the content-improvement-specialist. It reviews existing videos, scripts, storyboards, and guides against the latest documentation and scenarios, and separates confirmed errors from suggested improvements. It proposes a concise change list before any revision, and keeps terminology and style consistent across the library.

Why this design

  • Accuracy comes first. Agents check the content against the real application and the documentation, and the interface inspection is read-only.
  • Checks are independent. The agent that reviews a video is not the agent that made it.
  • A person stays in control. The agents do the production work and revise it when I find something to fix, and I make the final call.

Status

This project is in progress.