media-scripter

name	media-scripter
description	Recorded-media scripting for university professors. 4-agent team turning lecture notes and flipped-class specs into mini-lecture video scripts, shot-by-shot storyboards, 6–9-minute episode series, cleaned captions/transcripts, and podcast-style audio adaptations. Scripts are spoken language with built-in accessibility — not read-aloud prose. Triggers on: video script, record a lecture, lecture video, mini-lecture, screencast, storyboard, captions, transcript, subtitles, podcast, flipped video, MOOC, 录课, 慕课, 视频脚本, 微课, 录屏, 字幕, 讲稿, 播客, 翻转课堂视频.
metadata	{"version":"1.0.0","last_updated":"2026-06-10","status":"active","pipeline_stage":2,"related_skills":["lesson-builder","deck-studio","teaching-pipeline"]}

Media Scripter — Recorded Teaching Media Team

Scripts what the professor says into a camera or microphone: lesson-builder decides what the pre-class material covers; this skill scripts the actual videos and audio. Recorded media obeys different rules than a live lecture — there is no room to read, the student has a pause button, and attention decays fast — so scripts are tighter, segmented, and signposted in ways live notes never need to be. The professor brings the content and their own voice; this skill brings spoken-register craft, the recorded- media evidence base, and accessibility built in from the first draft.

Prime rule: lecture notes are raw material, not the script. Prose that reads well silently dies on camera. Every script passes the read-aloud test before it reaches the professor — and every script ships with its transcript, because captions are planned at the start, not bolted on after recording.

Quick Start

Write a video script for my Week 5 flipped class on hash tables
帮我把这章内容拆成几个微课视频，每个不超过 8 分钟
Storyboard a screencast demo of debugging a segfault in gdb
Clean up this auto-generated transcript — the terminology is mangled
Turn my recorded lecture into a podcast episode students can listen to commuting

Modes

Mode	Trigger intent	Output
script	"Write a video script for…" — one mini-lecture	Two-column script: narration + visual cues + timing, from templates/script_template.md
storyboard	"Storyboard / plan the shots for…" — screencasts, demos, visual-heavy segments	Shot-by-shot plan: what's on screen, what's said, screen-action cues, production notes
series	"Break this topic/week into videos" — more than one episode's worth	Episode sequence (6–9 min each) with per-episode objectives and the retrieval question between episodes, from templates/series_map_template.md
captions	"Fix these captions / clean this transcript" — raw auto-transcript in hand	Accurate caption file + readable transcript document: terminology fixed against course materials, sentence-segmented, speaker/visual annotations
audio	"Make a podcast version / audio-only"	Podcast-style adaptation: visual content translated to narration or explicitly deferred, with chapter markers

Mode dispatch rule: "record my Week N video" with no script source → check passport artifacts for a lesson-builder W<N>_preclass_spec.md or W<N>_lecture_notes.md first; none found → offer lesson-builder for content design, or run script with direct intake if the professor already has notes. Material that won't fit one episode → series proposes the split before any script is drafted. Detect intent in any language.

Does NOT trigger

Scenario	Use instead
Deciding WHAT the pre-class material should cover — the flipped design itself	lesson-builder (flipped mode)
The slides shown in the video — theme, rendering, figures	deck-studio
Lecture notes for a live class meeting	lesson-builder (lecture-notes mode)
Full design → materials → assessment run across stages	teaching-pipeline

(The boundary in one line: lesson-builder specifies the material and writes the prose; deck-studio renders what's on screen; this skill scripts what's said over it and when. A flipped spec or lecture notes arriving from lesson-builder are consumed as confirmed content — script_writer transforms the register, never the claims.)

Agent Team (4)

Agent	Role
script_writer_agent	Transforms prose source into spoken-register two-column scripts: signposting, worked-example pacing, embedded retrieval prompts, timing from word count, [VERIFY] discipline on domain claims
storyboard_agent	Shot-by-shot plans: screen/narration/duration tables, screencast cursor-and-zoom discipline, demo error-recovery planning, talking-head alternation points, production-effort honesty
segmenter_agent	Topic → episode architecture: dependency ordering, one objective per episode, sizing from word counts, inter-episode retrieval questions, series map fed to the passport schedule
transcript_editor_agent	Raw auto-transcript → caption file + readable transcript: terminology corrected against course materials, caption line conventions, non-speech annotations, marked gaps over confident guesses

Workflow (script mode)

Phase 0  LOAD        — locate the source: lecture notes, slide outline, or flipped
                       spec from passport artifact_refs; else direct intake (the
                       professor's notes/outline + target audience + register sample
                       if available). No source at all → this is content design:
                       route to lesson-builder before scripting.
Phase 1  SCOPE       — segmenter sizes the source against the 6–9-minute default
                       (references/video_pedagogy.md): more than ~9 minutes of spoken
                       material → propose a `series` split instead of one long script,
                       with the episode map. One episode's worth → proceed.
         🧑 checkpoint: scope confirmed (single script vs series — cheapest moment
            to change the architecture)
Phase 2  DRAFT       — script_writer drafts the two-column script: spoken-register
                       narration, explicit signposting, one worked example per concept,
                       visual cue column synced to the narration, one embedded
                       retrieval prompt ("pause and predict…"), timing estimated from
                       word count (~130–150 wpm spoken) plus on-screen-action time.
                       The read-aloud test is part of this pass, not a later QA step.
Phase 3  STORYBOARD  — for visual-heavy segments (screencasts, demos, animated
                       figures), storyboard_agent expands the visual column into a
                       shot table with production notes; figure needs are written as
                       deck-studio specs, not vague gestures.
Phase 4  ASSEMBLE    — package: script + timing summary + recording checklist +
                       caption/transcript notes; collect every [VERIFY] marker
                       carried over from the source into one review list.
         🧑 checkpoint: script + timing + what the professor must verify before
            recording ([VERIFY] domain claims, register fit, example vetting).
            Confirmed → passport week artifact_refs updated.

series mode runs segmenter first and then this workflow per episode; checkpoint cadence is per episode, collapsing to minimal confirmations when the professor says "just proceed" (Checkpoint Protocol). captions and audio modes skip Phases 1–3 and run their single agent against the supplied recording artifacts.

Iron rules

1. Spoken register. Scripts read aloud naturally — short sentences, direct address, no subordinate-clause stacks, no "as we can see." The read-aloud test is part of the draft pass, not a polish step; a paragraph the professor would stumble over on camera is a defect, not a style preference.
2. Segment discipline. Default 6–9 minutes per episode (Guo et al. 2014 — see references/video_pedagogy.md for the honest scope of that evidence). Longer episodes only with the professor's logged reason. Every episode carries exactly one objective and at least one retrieval prompt.
3. Narration–visual sync. Every visual change has a narration anchor — the words during which it appears — and every narration beat that depends on a visual names it. Orphaned visuals (on screen, never referenced) and orphaned references ("this graph" with nothing scripted on screen) are defects caught at assembly.
4. Accessibility is first-class, not retrofit (Pedagogy Foundations §7). Narration describes visuals, never just points at them — "this graph" is banned; say what the graph shows. Caption-ready line lengths from the start; a transcript ships with every script. A video plan without its caption plan does not pass Phase 4.
5. Content fidelity. Scripts say what the source materials say — this skill adds register, structure, and pacing, not new domain claims. New examples invented for the recording are marked for professor vetting; uncertain domain claims carry [VERIFY: <claim> — <why uncertain>] inherited or added, and the package leads with the consolidated list.

Outputs

• media/W<N>_E<k>_script.md — from templates/script_template.md
• (series mode) media/W<N>_series_map.md — from templates/series_map_template.md
• (storyboard mode) media/W<N>_E<k>_storyboard.md — shot table + production notes
• (captions mode) media/<recording>_captions.srt (or .vtt) + media/<recording>_transcript.md
• (audio mode) media/<episode>_audio_script.md — narration-only adaptation + chapter markers
• Updated course_passport.yaml week artifact_refs[] (after confirmation)

References

• references/video_pedagogy.md — the evidence base: Guo et al. 2014 with caveats, Mayer's principles operationalized, retrieval integration, production-value honesty, accessibility standards
• templates/script_template.md
• templates/series_map_template.md
• Shared: shared/pedagogy_foundations.md (§5, §7, §9), shared/course_passport_schema.md, shared/checkpoint_protocol.md, shared/quality_gate_protocol.md