Pptx2video

Convert an edited PPTX to video

This is a standalone, general-purpose PowerPoint-to-video package. It has no runtime dependency on paper2video, ResearchStudio-Reel, or ppt-master, and the complete CLI and renderer live in the installable pptx2video Python package. It is not limited to research papers or to a previously delivered video.pptx. Use any local PPTX as the visual source. Prefer canonical Author Notes for precise narration and marker timing. When canonical Notes are absent, read the same block grammar from Shape Alt Text. Never reuse rendered bundle artifacts. Reusing a pristine block-level TTS cache entry is safe only when its complete synthesis identity matches.

Entry point

Treat /pptx2video <input.pptx> as a direct invocation of this skill. Use the installed public CLI for all ordinary rendering:

pptx2video render <input.pptx> <new-output-directory> --resolution 1080p

Add --animation-pointer cursor|laser to follow native animation targets; it is independent of Spotlight and defaults to none.

Do not import or execute private Paper2Video or ResearchStudio scripts. If the CLI is unavailable, install the runtime from the public repository before continuing:

python -m pip install \
  'pptx2video[svg] @ git+https://github.com/ai-nuts/pptx2video.git'
python -m playwright install chromium
pptx2video doctor --svg

Authoring contract

For precise control, write a Notes block whose handle matches a top-level shape or group:

## [latency-card] latency-card
A new latency card appears. [[Spotlight] It reports lower latency.]

Alt Text uses the same single-block grammar, including both leading # characters. The semantic text after ] is optional and is not spoken:

## [latency-card] latency-card
A new latency card appears. [[Spotlight] It reports lower latency.]

The baseline hash, generated ordering provenance, and semantic profiles live inside the PPTX shape OOXML. The visible semantic after ] is the concise profile. PPTX generation or protocol writeback also persists a detailed profile. Conflict display reads those stored values and uses deterministic fallbacks for legacy decks, so it does not call an LLM. Native effect, target, trigger, and delay data stay in PowerPoint's p:timing tree. They do not appear in Notes or Alt Text. Older [handle] plus Script:, [ID] handle, and verbose [Paper2Video] blocks remain readable and migrate to this unified form on the next writeback. Keep Author Notes clean. Read references/editable_pptx.md for the hidden provenance contract.

The stable handle resolves the target directly. By default, successfully resolved canonical Notes blocks keep their current relative order as anchors. Only narrated targets missing from Notes are inserted around those anchors by current canvas geometry. The --narration-order-policy author-notes option is retained as an explicit compatibility alias and produces the same anchor-plus-insertion narration sequence. Animation Pane order independently controls native animation row order and local effect timing. In the rendered video, each Pane phase is placed in output-only pre-roll immediately before its narrated block, and the next phase waits for the preceding block to finish speaking. The only canonical script marker is Spotlight; its position determines an Edge word-boundary time or span. Legacy entrance markers remain readable for migration, but they never replace, create, or reorder a native animation.

Every delivered PPTX defaults to human-editable numbered click groups on every slide. Make each non-With Previous Pane row the On Click leader of its own top-level mainSeq group. Keep contiguous With Previous rows with that leader so they share its canvas badge. PowerPoint derives the badge from this topology; it is not a stored animation ID. Accept the rewrite only when effect targets, presets, durations, delays, Pane start/end times, row order, and simultaneous groups remain identical. Fail closed when the schedule cannot be preserved. Use --click-group-policy preserve only when the source's automatic click-group topology must remain unchanged.

For a PPT-native workflow without canonical Notes, put an explicit script in the shape's Alt Text using the same grammar, and use the native Animation Pane for effects:

## [latency-card] latency-card
A new latency card appears and reports lower latency.

Read references/animations.md before changing effect names or resolving a protocol conflict.

Authority and conflict rules

Apply this precedence:

1. An explicitly selected user script.json owns narration for that render.
2. Canonical Author Notes own handles. For narration, compare Notes and Alt Text with the last system-synchronized script hash stored in shape OOXML. A change on only one surface wins; if both changed differently, Notes wins and the authority report records the conflict.
3. A canonical Alt Text block provides narration when Notes are absent or when its narration alone differs from the stored baseline. A new animated target with only plain pre-protocol Alt Text uses that text once as its initial script and is normalized to the unified block grammar.
4. Narration order defaults to incremental geometry insertion. Every successfully resolved canonical Notes block is an anchor and keeps its relative Notes order. Only narrated targets absent from Notes are inserted around those anchors. At least 35 percent vertical overlap forms horizontal bands read top-to-bottom. Within each band, at least 50 percent horizontal overlap forms vertical lanes read left-to-right, and every lane is completed top-to-bottom before advancing right. Moving an existing anchored shape does not reorder established narration; move its Notes block when that order must change. Silent targets without geometry remain deterministic at the end. The explicit author-notes policy remains available for compatibility and uses the same Notes-anchor plus geometry-insertion sequence. Both policies fail closed when a newly inserted narrated block has no usable top-level geometry. After resolving Notes, Alt Text, an explicit script.json, or requested changed-element regeneration, but before TTS or output-bundle creation, compute two deterministic entrance sequences: a pure full-slide band/lane geometry order and native Animation Pane order. Number every eligible element again in each sequence; do not present the Notes-anchor narration sequence as geometry order. A conflict exists only when a relative inversion involves a newly inserted, Notes-missing narrated entrance target. This candidate filter does not geometrically reorder established Notes anchors. Ignore silent decoration, emphasis-only targets, and simultaneous groups. Emit animation_order_authority.json containing only slides with such conflicts, including exact Pane-row move guidance. Slides without conflicts remain silent and require no decision.
5. Animation Pane is the sole authority for ordinary native animation type, row order, On Click, With Previous, After Previous, duration, and delay. The video scheduler preserves those local relationships while adding narration release gates between Pane phases. A script can supplement the native effects only with Spotlight.
6. Shape OOXML stores orderSource and canonical orderIndex beside the script hash. A no-conflict render writes author_notes for established Notes anchors and geometry for newly inserted blocks, then rewrites contiguous indices for the effective sequence. A confirmed Pane choice writes animation_pane for that selected conflict slide's effective entrance sequence.
7. The PowerPoint canvas owns all visible pixels and geometry.

All 22 registered native entrance presets have explicit MP4 strategies. Complex presets use named deterministic approximations and never silently become Fade. Unknown entrance preset tuples and native animation classes outside the supported entrance/emphasis contract fail explicitly.

When Notes and Alt Text both changed differently, Notes wins and the delivered PPTX Alt Text is refreshed. When a legacy script entrance marker conflicts with a native row, the native row wins and the conflict report records the ignored marker. Without a native row, a legacy entrance marker is ignored rather than inventing an animation. If a shape was deleted, remove its stale Notes block and record it in removed_stale_notes; never rebind it by position or by a remaining Animation Pane row. Ambiguous live handles still fail closed.

One-command render

Install the public runtime and verify both core and SVG dependencies before rendering:

python -m pip install \
  'pptx2video[svg] @ git+https://github.com/ai-nuts/pptx2video.git'
python -m playwright install chromium
pptx2video doctor --svg

Then invoke the local CLI:

pptx2video render \
  path/to/edited.pptx \
  path/to/new_video_bundle \
  --resolution 1080p

--narration-order-policy geometry is the default incremental policy. It preserves the current canonical Notes block sequence and uses geometry only to place narrated targets missing from Notes. --narration-order-policy author-notes remains an explicit compatibility selection.

The default --click-group-policy normalize processes every slide, not only the first slide. It produces one numbered badge per sequential phase while keeping With Previous rows in the same numbered group. preserve is the explicit opt-out.

The default --animation-order-policy auto is interactive only when stdin is a TTY. It prompts only for conflict slides, in ascending slide index, and waits for each decision before showing the next conflict slide. For example, resolve slide 2 before displaying slide 4. In non-interactive or API use, a detected conflict writes the report and exits with code 3 before creating the output directory.

When Codex or another skill presents that report in conversation, show exactly one conflict slide at a time and wait for the user's choice before showing the next. Use the report's deterministic text or ASCII canvas, draw each element's OOXML bounding box rather than only its center, and show three vertically stacked views that approximate the PPTX canvas's actual aspect ratio after accounting for monospace character-cell proportions. Fit all dimensions to the detected terminal width rather than clipping a canvas. Use identical outlines in all three views. Put only the geometry rank number inside each box in Geometry Order, only the Pane rank number inside the same box in Animation Pane Order, and only its stable identity letter such as A, B, or C in Element Identity Map. Do not use combined labels such as G1/P2. Identity letters follow natural sort of stable shape_id, independently of both order sequences, so they carry no ordering meaning. Give every canvas its own key-to-semantic legend titled exactly LEGEND. Beside Geometry Order, map each geometry rank number to that element's semantic description. Beside Animation Pane Order, map each Pane rank number to that element's semantic description. Beside Element Identity Map, map each identity letter to that element's semantic description, so a user can state an arbitrary preference such as C -> A -> D -> B. On a wide terminal, place each legend to the right of its corresponding canvas. On a narrower terminal, place it immediately below that canvas before presenting the next view. The same selected profile applies to all three legends: default to concise, or use the stored detailed profile through --semantic-profile detailed. Below 26 columns, omit the bbox plots but retain three labeled textual views, each with its own numeric-rank or identity-letter semantic mapping. Thin one-line elements must remain visibly wide and thin. Do not use an LLM to infer either sequence, the identity map, the layout, or the semantic text at conflict-display time, and do not list all conflict slides at once. A Pane choice applies only to the selected conflict slide; it serializes that slide's eligible narrated entrance blocks in Pane order and synchronizes delivered Notes and OOXML provenance. Other conflict slides keep their own decisions, and no-conflict slides retain Notes-anchor plus inserted-geometry provenance. If the user chooses reading-order on any slide, stop and present the exact Pane move guidance from the report; do not silently render a video whose schedule differs from the delivered PPTX.

Parse a user-selected Identity order with deterministic identity_order.v1. For a conflict page whose Identity Map is A through G, CABGDEF means exactly C -> A -> B -> G -> D -> E -> F. Treat input case-insensitively and canonicalize it to uppercase. Require a complete permutation of the current page's displayed identities, rejecting missing, duplicate, unknown, empty, or invalid labels. Compact undelimited input is valid while every label is one letter. If a page extends past Z to AA, require separators such as A,B,...,AA. In a TTY, accept the order directly for the displayed page. For API or non-TTY use, accept --animation-order-sequence CABGDEF for one conflict page, or repeat --animation-order-sequence SLIDE=ORDER for every conflict page. If the sequence exactly matches the current Pane, record it as explicit Pane confirmation and render. If it differs, persist its Identity and shape-ID orders in the decision report and stop for a source-PPTX Pane revision. Never render an order that disagrees with the delivered PPTX.

Read references/cli.md for installation and arbitrary-path usage. Add --script-json path/to/edited-script.json only when a user-edited external script should override PPTX narration. This command must:

1. Normalize unified Alt Text blocks from authoritative Notes when present.
2. Generate handles and unified Alt Text blocks for new animation targets.
3. Backfill canonical Notes when the source has only Alt Text or native rows.
4. Extract narration using user script, Notes, then Alt Text precedence, and apply the selected narration-order policy independently of script authority.
5. Normalize numbered click groups across every slide after change detection and protocol writeback. Verify that the renderer's Pane clock is unchanged and write assets/meta/reports/click_group_normalization.json.
6. Materialize Edge TTS and word timings per narration block, reusing only exact cache hits. During output assembly, prepend the corresponding Animation Pane phase as silence before each narrated block, shift every Edge word boundary by the same amount, then concatenate the blocks into slide audio. Generate deterministic silent audio for a native-only silent slide. Preserve each logical section ID in metadata while resolving its portable physical MP3 through manifest.json.
7. Build the narration-serialized schedule: a narrated entrance completes before its first spoken word, and the following Pane phase starts only after the preceding narrated block ends. Preserve native simultaneous groups.
8. Align every subtitle cue to its actual first and last Edge TTS word boundaries. Never use proportional timing in a final render.
9. Render PPTX pixels, animations, audio, spotlight, and bottom-band subtitles. Read references/render_video.md for frame-source, attention-overlay, ffmpeg, performance, and debugging details.
10. Write timeline.json, subtitle-alignment evidence, mapping reports, and strict QA evidence.

Do not pass --prebuilt-audio-dir for a final render. Do not pass --no-qa.

The persistent TTS cache is content addressed per block. Its invalidation key is the SHA-256 of a canonical JSON identity containing normalized narration SHA-256, voice, rate, pitch, provider, provider version, and cache schema version. A timing-aware render accepts a hit only when audio, metadata, and word-boundary evidence all match their hashes and cache key. Editing one block therefore regenerates that block only; changing any synthesis setting or provider version invalidates every affected key. Animation pre-roll and tail padding happen after cache materialization and never contaminate the pristine cache entry.

For a change-aware rerender, ordinary users choose whether to keep existing narration or regenerate only changed elements. The latter requires a previous PPTX baseline and an OpenAI API key:

pptx2video render edited.pptx new-bundle \
  --baseline-pptx previous-video.pptx \
  --narration-mode regenerate

Regenerated scripts are written back into both Author Notes and unified Alt Text blocks in the delivered PPTX. script.json remains an Advanced override, not the normal editing surface.

Ordinary animated PPTX

If the deck does not yet contain canonical Notes and Alt Text, bootstrap a copy from an existing narration script, then edit that copy:

pptx2video bootstrap path/to/animated-source.pptx \
  --script-json path/to/script.json \
  --output path/to/editable-video.pptx \
  --report path/to/bootstrap-report.json

Completion gate

Require these deliverables:

new_video_bundle/
  video.mp4
  video_no_subtitles.mp4
  video.pptx
  manifest.json
  assets/audio/
  assets/captions/
  assets/meta/timeline.json
  assets/meta/reports/author_notes_authority.json
  assets/meta/reports/animation_order_authority.json
  assets/meta/reports/click_group_normalization.json
  assets/meta/reports/subtitle_timing_alignment.json
  assets/meta/reports/video_qa_report.json

Confirm the subtitle timing report has status: word_aligned, then confirm video_qa_report.json has passed: true, error: 0, and warning: 0. Do not claim completion until the strict renderer exits 0.

macOS local use

Read references/macos.md. Install the same Python package, install Playwright Chromium, verify native and SVG dependencies with pptx2video doctor --svg, and invoke the standalone CLI from Terminal or an external launcher.