Produce still-slide visuals, generate or ingest narration audio, assemble simple narrated presentation videos, validate exports, and package final delivery.
Instrucciones de origen · Vista previa de solo lectura
name
slide-video-producer
description
Produce still-slide visuals, generate or ingest narration audio, assemble simple narrated presentation videos, validate exports, and package final delivery.
Slide Video Producer
Use this skill to produce the final narrated presentation video after the narration script and slide storyboard pass full review.
This role intentionally merges slide visual design and final video assembly because the format is simple: approved slide images plus approved narration audio. It should not become a motion-graphics production role by default.
Expected Inputs
approved presentation-brief.md
research-notes.md when present
source-material-inventory.md when present
passed narration-review-report.md
reviewer-passed narration-script.md
reviewer-passed slide-storyboard.md
user-provided images, screenshots, charts, logos, or source visuals when available
user-supplied narration audio when provided
language, voice, subtitle, aspect ratio, resolution, and export constraints
approved visual mood/style contract
approved audio persona/performance contract
approved narrator identity and voice-consistency contract
approved prompt-level cue style for generated speech
Produced Artifacts
media-resource-index.md
slide-video-production-plan.md
final slide images under slides/
slide-video-production-log.md
voiceover-package.md
audio-generation-log.md when generated speech or imported audio processing is used
image-generation-log.md when generated or edited slide images are used
Use it as the shared reference for slide simplicity, source discipline, narration-slide mapping, required generate_image / edit_image slide creation, embedded slide content, generate_speech-only narration generation, audio-led assembly, optional subtitles, and final validation.
The only tools allowed for final slide image creation are generate_image and edit_image.
Image generation and image editing calls require a 60-second cooldown between calls.
The only generated narration tool allowed is generate_speech after checking list_audio_models.
Prefer create_video_from_image_and_audio and concatenate_videos for still-slide video assembly.
Treat speak, Python/PIL-only final slide creation, HTML/SVG-only final slide creation, and script-only final slide creation as disallowed routes.
Workflow
Step 1 - Confirm the approved package
Before production, confirm:
narration review decision is Pass
final narration script exists
slide storyboard exists
narration-to-slide mapping is explicit
required embedded slide content is specified for each slide, including exact on-image wording or labels when needed
language, aspect ratio, resolution, and subtitle requirements are clear
approved visual mood/style contract is clear
approved audio persona/performance contract is clear
approved narrator identity and voice-consistency contract is clear
approved prompt-level cue style is clear
If the script has not passed full review, route the package back to narration_script_reviewer or presentation_director.
If the mapping is unclear, route the package back to presentation_director.
Step 2 - Create the production package skeleton
Before any generate_image, edit_image, generate_speech, subtitle, segment-video, or final assembly call, create the package files from templates or minimal placeholders:
media-resource-index.md
slide-video-production-plan.md
image-generation-log.md
voiceover-package.md
audio-generation-log.md
slide-video-production-log.md
video-assembly-package.md
final-delivery-report.md
This is a hard gate. Do not begin media generation with these files missing. Keep them as live records; do not reconstruct them only at the end.
In media-resource-index.md, inventory sources, generated resources, logs, and final outputs:
user-provided images, screenshots, charts, diagrams, logos, and documents
source images extracted from supplied material
generated or edited illustration assets
final generated or edited slide images
generated or imported audio clips
subtitles when present
audio-generation-log.md
image-generation-log.md
production logs, assembly package, delivery report, and final exports
For each entry, record resource id, type, tool or method, path or URL, origin, related slide id, related clip id, what it contains, derivation lineage, rights or sensitivity notes when known, status, and final-use status.
Step 3 - Write slide-video-production-plan.md
For each slide, plan:
linked narration segment ids
visual type: text slide, process infographic, illustrated presentation slide, pictorial storyboard slide, diagram, chart, image-led slide, screenshot-led slide, comparison, timeline, or summary
title or short headline
main message
required embedded slide content: exact on-image wording when needed, labels, diagram content, chart content, source visuals, and visual structure the final slide image must contain
visual mood/style target for the slide, normally light, friendly, relaxed, clear, and audience-approachable
visual standard / engagement intent, normally polished, high-quality, content-bearing, and visually engaging without becoming busy
image-only or low-text rationale when no written wording is planned
layout notes
visual asset or diagram needs
source basis
readability risks
visual fit for the linked narration audio
For audio and assembly, plan:
narration source: generated speech or user-supplied audio
selected narrator identity for the whole video
audio persona/performance target, normally warm, relaxed, conversational, and engaging
prompt-level cue plan for natural generated delivery
one clip per narration segment by default
slide-to-clip mapping
subtitle mode
transition mode: cuts by default, light fades only when useful
Do not add new explanatory claims to slides.
If a slide needs new information to work, route that issue to presentation_director.
If the approved mapping would make a slide visually stale, overloaded, or too thin for the narration, route a storyboard issue upstream instead of silently changing the script.
If a storyboard slide lacks required embedded slide content, route the package upstream before production.
Step 4 - Produce still slides
Create one final slide image per storyboard slide unless the approved storyboard explicitly calls for a different split.
Required production route:
Final slide images must be created with generate_image or edit_image.
Use generate_image for new presentation/infographic slides, concept visuals, explanatory diagrams, timelines, process slides, comparison slides, summary slides, or slide compositions created from approved content.
Use edit_image when starting from user-provided screenshots, charts, source images, prior generated slides, or existing visual drafts.
Do not use Python/PIL-only, HTML/SVG-only, presentation-tool-only, or script-only rendering as the route for final slide visuals.
Local scripts may prepare source assets, measure dimensions, convert formats, inspect files, post-process already generated/edited outputs, assemble video, or create validation evidence, but they must not replace generate_image or edit_image for final slide image creation.
The generated or edited output must already be the finished content-bearing slide image. Do not generate a decorative background and add the real slide content later with scripts, PIL, HTML/SVG, presentation tools, ffmpeg overlays, or other post-generation overlay mechanisms.
Optional narration subtitles are allowed when requested or useful for accessibility, but they must not be used to compensate for missing embedded slide content.
Generated image and edit rules:
Run image generation and editing serially.
Call exactly one generate_image or edit_image request at a time.
Wait for the result, failure, or timeout, run the minimum visual acceptance check when an image is present, and immediately update image-generation-log.md with that attempt and decision.
Immediately run sleep 60 before any further generate_image or edit_image call.
Do not dispatch image calls concurrently, in the background, or as a parallel batch.
Record every image attempt in image-generation-log.md with full prompt text, full generation/edit config JSON, and 60-second cooldown evidence. Update this file after each accepted, rejected, failed, timed-out, regenerated, or edited result before the cooldown and before the next image call; do not reconstruct it only at the end.
Record generated or edited asset lineage in media-resource-index.md.
Do not represent generated illustrative visuals as source evidence.
Reject and regenerate or edit any output that omits required embedded slide content, corrupts planned wording or labels, invents unsupported wording, or loses the slide's intended explanation.
Reject and regenerate or edit any output that violates the approved visual mood/style contract, including outputs that are too dark, solemn, ominous, poster-like, heavy, or unfriendly when the approved style is relaxed and light.
Reject and regenerate any output that has an obvious visual failure: low-quality, static, generic, draft-like, poorly composed, unreadable, or clearly not engaging enough for the approved viewer experience.
If the base candidate is globally weak, do not try to polish it with edit_image. Regenerate from scratch with generate_image using the approved storyboard and a stronger visual storytelling prompt.
For every retry, name the specific defect being fixed and preserve what already worked. Do not change the slide's visual type, composition model, or level of illustration unless that was the actual defect.
If a pictorial panel slide is close but has one unwanted visual element, prefer edit_image or a targeted regeneration prompt that preserves the pictorial panel structure. Do not rewrite it as a minimal icon/card slide unless the approved storyboard calls for an icon-based process infographic.
Avoid vague retry instructions like "clean infographic" by themselves; they can collapse a rich illustrated slide into flat icons. Specify the intended type, for example illustrated presentation slide with three pictorial cards or process infographic with simple icons.
A retry prompt should include a preservation clause, for example: Preserve the previous attempt's light watercolor style, three pictorial cards, human scene illustrations, title hierarchy, and friendly mood. Only remove the unwanted oil press object and keep all required text accurate.
If repeated generation/editing cannot produce the required embedded content correctly, block and route upstream instead of silently delivering an incomplete slide image.
Do not run a separate detailed visual-critique loop by default. If the user visually approves the generated slides and no required-content or obvious visual failure is present, record that approval in image-generation-log.md / slide-video-production-log.md and proceed. A user may waive detailed visual critique, but that never waives live logs, resource indexing, or final package records unless explicitly stated.
Slide rules:
one primary message per slide
readable at target resolution
high-quality, polished, and engaging enough to hold viewer attention while the narration explains the point
no dense paragraphs
no unnecessary decorative complexity
visual mood should match the approved style contract; default to light, relaxed, friendly, clear, and audience-approachable
avoid dark cinematic, solemn, ominous, black/gold poster-like, or heavy visual styles unless explicitly approved
every final slide image is a complete content-bearing presentation or infographic image, not a plain background awaiting later overlays
no accepted slide should have an obvious low-quality, draft-like, static-placeholder, or generic first-pass failure
where text, labels, or diagram wording are planned, they should be readable and faithful to the approved storyboard
embedded slide content should support the narration, not duplicate it line for line
charts, diagrams, and screenshots must be accurate to the source basis
the slide must work for all linked narration segment ids without becoming stale or overloaded
Step 5 - Create voiceover-package.md
Record:
narration source: generated speech or user-supplied audio
voice choice or imported audio path
selected narrator identity for the whole video
approved audio persona and performance target
approved narration text per clip
full speech prompt text planned for generate_speech
prompt-level audible performance cues, when used
subtitle text without non-spoken cues
linked segment and slide id or approved slide sequence
performance directions
approval state
Do not change approved narration wording.
When generate_speech is used, the speech prompt should normally include concise non-spoken bracketed performance cues around the approved narration, such as [warm, conversational], [short pause], [gentle emphasis], or [slightly brighter], to make the delivery more natural and less stiff.
Use only cues that can affect audible delivery.
Do not add visual directions, new claims, new explanation content, or cue text that should appear in subtitles.
Use one consistent narrator voice/persona for all generated clips unless the approved brief explicitly requires multiple voices.
Changing local tone should be done through prompt-level cues, not by changing voice identity clip by clip.
If a line is too long, awkward, or cannot be generated naturally from the approved script, route it to presentation_director and expect another full narration review before continuing.
Do not create new mid-clip slide changes that were not approved in the storyboard.
Step 6 - Generate or process narration
If generated narration is needed, first call list_audio_models, select a suitable model or voice, build a full generation config from the live model schema, and generate one speech clip per approved segment with generate_speech only, unless the user supplied approved audio.
Do not use speak.
Do not rely only on broad config fields such as voice, accent, tone, or style.
Select one narrator identity for the whole video before generating the first clip, then keep that same voice/speaker identity, model route, language/accent target, and broad generation config for every generated clip unless the approved brief explicitly says otherwise.
Build the actual prompt sent to generate_speech from the approved narration plus concise audible bracketed performance cues for warmth, emphasis, or pauses unless cues are unsupported by the selected model or explicitly inappropriate for the approved style.
Keep config and prompt responsibilities separate:
speech prompt text: approved narration plus local audible cue tags.
audio log: records both the full config and the full speech prompt.
Good style_instructions example:
Warm, relaxed Mandarin explainer narration. Conversational, clear, approachable, not sermon-like or dramatic.
Bad style_instructions example:
Read the Chinese transcript completely. Bracketed cue tags such as [warm, conversational], [short pause], and [gentle emphasis] are delivery directions, not words to speak.
Why the bad example is wrong:
it puts cue-parsing instructions into config instead of the speech prompt
it duplicates prompt mechanics instead of defining the broad voice
it encourages future clips to mix prompt text and config responsibilities
Example single-speaker prompt for a relaxed narrated presentation:
Opening a topic: [warm, conversational] 今天我们先从一个简单问题开始。
Creating a natural pause: 这个细节很小， [short pause] 但它解释了后面的转折。
Softening a serious point: [gentle, reassuring] 这里不是为了责备门徒，而是让我们看见人的有限。
Guiding a transition: [slightly brighter] 明白这个背景之后，我们再看耶稣接下来的提醒。
Emphasizing a takeaway: 所以重点不是他们多失败， [gentle emphasis] 而是耶稣仍然主动预备他们。
Cue misuse examples:
Do not use visual-only cues: [camera zooms in], [slide changes], [show olive trees].
Do not overload every sentence with cues; one to three audible cues per clip is normally enough.
Do not use cues to add new meaning that is absent from the approved narration.
Do not use different voices for different slides just to make the audio feel varied. Keep one narrator and use cues for local delivery changes.
Do not put cue examples or cue-handling explanations into style_instructions; keep those in the speech prompt and logs.
Speech generation is serial-only:
call exactly one generate_speech request at a time
wait for the result, failure, or timeout
inspect the output when present and immediately update audio-generation-log.md, including approved narration text, full speech prompt text sent to the tool, prompt-level performance cues, selected narrator identity, full generation config JSON, model/tool id, output path, voice-consistency check, inspection result, and any rejection/regeneration notes
update audio-generation-log.md after each accepted, rejected, failed, timed-out, or regenerated result before the next speech call; do not reconstruct it only at the end
do not dispatch speech calls concurrently, in the background, or as a parallel batch
reject and regenerate any clip that uses a different narrator voice, speaker identity, language/accent target, or noticeably different persona from the approved voice package
reject and regenerate audio that sounds too serious, stiff, old-fashioned, stern, sermon-like, monotonous, or boring when the approved audio persona is relaxed and conversational
If a model truncates or mishandles a long approved segment, split that segment into shorter subclips without changing the approved wording. Generate the subclips serially with the same narrator identity and config, log every subclip attempt in audio-generation-log.md, concatenate the subclips into the approved segment clip, inspect the combined clip for completeness and boundary smoothness, and record the subclip lineage in voiceover-package.md and media-resource-index.md.
If using user-supplied audio, split or map the audio to segment ids only when the mapping is clear.
Step 7 - Create optional subtitles
Create subtitles.srt only when requested, useful for accessibility, or required by the delivery context.
Subtitles should:
match the approved narration
remain readable
avoid covering important slide content
default to a conservative bottom-safe style for 1080p exports: small enough to support listening without dominating the slide, normally font_size about 18-20, bottom-center alignment=2, low margin_v about 10-16, thin outline, and a modest semi-transparent background only if it improves readability
not be placed high on the slide by default; keep them near the lower safe area unless the slide itself has essential lower-third content, in which case choose the least intrusive position and document it
avoid large multi-line blocks that cover central slide titles, diagrams, faces, or prayer/template text; prefer shorter SRT cue chunks when a sentence would create a wide or tall subtitle block
be checked on at least one early-frame sample after the first subtitle export; if the subtitle covers main slide content or feels visually dominant, immediately create a corrected smaller/lower subtitle version before delivery and make that corrected version the final export
Step 8 - Assemble the still-slide video
Use the video/audio tool route by default.
Default assembly:
create one video segment for each approved slide/audio pair with create_video_from_image_and_audio
concatenate accepted segment videos in approved slide/storyboard order with concatenate_videos
use cuts by default
use light fades only when useful and safe
preserve target aspect ratio and resolution
Do not use raw ffmpeg stream-copy concat of independently encoded segments as the default final route because it can create small audio discontinuities at segment boundaries.
If the video/audio tools are unavailable and ffmpeg is necessary, avoid a final -c copy concat unless the exported file passes boundary audio validation; otherwise re-encode the final concat or choose another documented route.
Do not create a separate continuous concatenated audio track as a routine workaround.
Only use a continuous-audio mux route when segment-video concatenation fails validation and the reason is documented in video-assembly-package.md and final-delivery-report.md.
Do not add pan, zoom, animated text, motion graphics, music beds, or sound effects unless explicitly approved.
Step 9 - Validate and package
Write slide-video-production-log.md, video-assembly-package.md, and final-delivery-report.md.
Validate:
final file opens successfully
correct aspect ratio and resolution
no black frames or missing slides
on-image wording, labels, diagrams, charts, and source images are readable at export resolution
every final slide image contains the required embedded slide content from the reviewer-passed storyboard
every final slide image passes the minimum visual acceptance check or has explicit user visual approval recorded; no accepted slide has an obvious low-quality, draft-like, generic, or visually dull failure
no post-generation text, label, diagram, callout, or explanatory slide-content overlay was used to compensate for an incomplete generated/edited slide image
final slide images are generate_image or edit_image outputs, with lineage recorded in media-resource-index.md
image-generation log includes full prompts, full configs, inspection decisions, and cooldown evidence for every image attempt, recorded immediately after each result rather than reconstructed only at the end
audio-generation log includes approved narration text, full speech prompt text, prompt-level performance cues, selected narrator identity, voice-consistency checks, inspection decisions, and full configs for every generated speech attempt, recorded immediately after each result rather than reconstructed only at the end
visual mood/style matches the approved contract and is not too dark, solemn, or heavy unless explicitly approved
generated narration keeps one approved narrator voice/persona across clips, matches the approved audio persona, and is not too serious, stern, sermon-like, or boring unless explicitly approved
slide visuals remain source-supported or clearly marked as generated/representative
audio is audible and complete
no audio drops, skipped syllables, or silence glitches occur at segment boundaries
each slide matches the narration
segment videos were created from the correct slide/audio pairs
subtitles align and are readable when present
no unapproved narration or slide content appears
final export path and validation status are recorded
Before final handoff, run a package-completeness check. If any required production package artifact is missing or empty, write or repair it before delivery:
media-resource-index.md
slide-video-production-plan.md
image-generation-log.md
voiceover-package.md
audio-generation-log.md
slide-video-production-log.md
video-assembly-package.md
final-delivery-report.md
Send the final package back to presentation_director or the user with absolute paths and known limitations.
Routing Rules
Script Wording Or Natural Speech Issue -> presentation_director
