You are a PIPELINE EXECUTOR. Your job is: prepare a reproducible prompt and
optional reference images/videos -> call Oracle's Gemini Create Video browser tool ->
persist the result payload and video artifact -> verify the local MP4/WebM/MOV
with ffprobe -> summarize the exact paths. For multi-clip stories, you also own
the continuity chain, the stitch, and the visual sanity check.
files accepts multiple reference images (PNG tested) AND previously
generated MP4s as continuity references in the same call — Gemini uploads all of
them through the same attachment path before submit.~/.oracle/config.json
(browser.remoteHost + browser.remoteToken). The saved video lands LOCALLY at
~/.oracle/sessions/{slug}/videos/video.mp4; logs at
~/.oracle/sessions/{slug}/output.log and
~/.oracle/sessions/{slug}/models/gemini-create-video.log (there is no log.txt).mcp__oracle__create_video.[{slug}] on its own first line, then a blank line.mcp__oracle__sessions before deciding.Agree.This skill uses Gemini Create Video through Oracle. Do not manage ChatGPT's Intelligence pill here.
Critical rules:
browserThinkingTime for mcp__oracle__create_video.Instant and Pro Extended for Gemini
video generation; those labels are ChatGPT-specific.mcp__oracle__create_video tool.Choose a slug: 3-5 hyphenated words, e.g. hero-loop-video-proof.
Create a run folder:
RUN_DIR="runs/$(date +%Y-%m-%d)-{slug}"
mkdir -p "$RUN_DIR"/{receipts,prompts,results,videos}
$RUN_DIR/prompts/create-video.md:[{slug}]
{video generation prompt — include "full 10 second" explicitly for max length}
Audio constraints: no music, no score, no singing, no lyrics, no dialogue, no voiceover, no spoken words, no mouth-synced speech. Prefer a silent clip; if audio is unavoidable, use only subtle non-musical ambient sound.
Also end visual prompts with: No text, no captions, no readable words.
mcp__oracle__create_video({
prompt: "[{slug}]\n\n{video generation prompt}\n\nAudio constraints: no music, no score, no singing, no lyrics, no dialogue, no voiceover, no spoken words, no mouth-synced speech. Prefer a silent clip; if audio is unavoidable, use only subtle non-musical ambient sound.",
slug: "{slug}",
files: []
})
For image-conditioned video, pass absolute paths (order matters — describe them as "sequential visual references" and reference them in order in the prompt):
mcp__oracle__create_video({
prompt: "[{slug}]\n\nUse the attached images as sequential visual references. {prompt}\n\nAudio constraints: ...",
slug: "{slug}",
files: ["/absolute/path/reference-01.png", "/absolute/path/reference-02.png"]
})
For continuity from a prior clip, put the previous MP4 FIRST in files and say
"Use the attached previous video as continuity reference for the camera language,
color grade, and tone" (see guardrail ladder before using identity language).
Optional debugging only: browserKeepBrowser: true, browserConversationUrl: "...".
CLI recovery shape (same semantics; useful for batches and when MCP is unavailable):
oracle --engine browser --model gemini-create-video --slug {slug} \
--prompt "$(cat "$RUN_DIR/prompts/create-video.md")" \
--file /abs/ref-01.png --file /abs/ref-02.png \
--verbose --wait --force
$RUN_DIR/results/create-video-result.json$RUN_DIR/results/create-video-output.log (tail of ~/.oracle/sessions/{slug}/output.log)videoPaths[] entry (typically
~/.oracle/sessions/{slug}/videos/video.mp4) into $RUN_DIR/videos/ with a
descriptive name.ffprobe -v error -show_entries format=duration,size:stream=codec_type,codec_name,width,height,duration \
-of json "$RUN_DIR/videos/{name}.mp4" > "$RUN_DIR/receipts/ffprobe.{name}.json"
Duration must be greater than zero and should match what was requested (e.g.
10.000000 for a full-length clip). Include codec, dimensions, duration, and size
in the final summary.
Healthy run log markers: Gemini Create video tool selected.,
Attachment upload complete in Gemini: <each file>, files=N,
Gemini video still generating... (generating), Saved 1 generated video to ....
Proven shape: 10 story images -> four 10s clips -> 40s straight concat -> 60s paced edit.
runs/YYYY-MM-DD-{story-slug}/{prompts,results,receipts,videos} plus a
manifest.md mapping each segment to its images, continuity reference, prompt
file, and Oracle session slug. Write ALL per-clip prompt files up front.~/.oracle/sessions/{clip-slug}/videos/video.mp4 ->
$RUN_DIR/videos/NN-{beat-name}.mp4, save the output.log tail, write the ffprobe
receipt, and only then start the next clip with this MP4 as its first files entry.cd "$RUN_DIR"
cat > concat.txt <<'EOF'
file 'videos/01-....mp4'
file 'videos/02-....mp4'
...
EOF
ffmpeg -hide_banner -y -f concat -safe 0 -i concat.txt -c copy {story}-40s.mp4
# Optional ~60s paced edit (1.5x duration; re-encode):
ffmpeg -hide_banner -y -i {story}-40s.mp4 \
-filter_complex '[0:v]setpts=1.5*PTS[v];[0:a]atempo=0.6666667[a]' \
-map '[v]' -map '[a]' {story}-60s.mp4
For final deliverables, strip generated audio in the stitch (-an) unless the
user explicitly asks to preserve it; add music/narration/sound design only after
the picture edit is locked, as one continuous pass.
5. Visual sanity receipt — contact sheet, then actually look at it:
ffmpeg -hide_banner -y -i {story}-60s.mp4 -vf 'fps=1/5,scale=320:-1,tile=3x4' \
-frames:v 1 receipts/contact-sheet-60s.jpg
SUMMARY.md in the run folder: outputs, per-clip sessions/images/refs,
ffprobe receipts, and any guardrail notes.Gemini may refuse with text like "I can't make content that appears to show specific people." This fires probabilistically on identity-continuity prompts — the same shape (prior MP4 + person images + "same man") can succeed for several clips and then fail. When it fires, do NOT retry the same prompt. Step down:
Record which rung succeeded in the run notes so the stitch summary explains any continuity seams.
(generating) polls. (complete-without-video) polls
mean Gemini answered with text but produced no <video> — almost always a
guardrail/refusal. Do not wait out the 20-minute cap: inspect the live run
target, read the visible response text, then cancel and apply the ladder.scripts/browser-tools.ts eval and
/json/activate only reach the active tab, which is often NOT the isolated run
target. Instead: curl http://127.0.0.1:<chrome-debug-port>/json/list on the tv,
identify the Gemini page target whose body contains this run's prompt text, then
connect straight to its webSocketDebuggerUrl with Node's built-in WebSocket
(Node >= 22; no ws package needed) and send Runtime.evaluate to read
user-query / model-response counts and visible text. A genuinely submitted run
shows 1 user-query, 1 model-response, and text like "I'm generating your video.
This could take a few minutes".f1d6231b by requiring a real submitted
turn/response or prompt disappearance). If you see it again, treat it as a send
failure: cancel, fix/report, rerun.ps -axo pid,command | rg 'gemini-create-video').curl -X POST -H "Authorization: Bearer $TOKEN" http://{remoteHost}/runs/{runId}/cancel (find the runId via GET /runs).GET /runs and check .active is empty.
The /health activeRuns counter can leak after cancellations and stay nonzero
while no run is actually active.| Symptom | Action |
|---|---|
create-video-tool-not-found |
Gemini menu DOM changed. Re-probe the remote Gemini tab and update selectors before retrying. |
create-video-chip-missing |
The menu click did not activate the Videos composer chip. Retry once after inspecting the composer DOM. |
Polls show (complete-without-video) |
Guardrail/refusal in the response text. Inspect the run target via CDP, read the message, cancel both sides, apply the guardrail ladder. |
| "can't make content that appears to show specific people" | Apply the guardrail ladder; the proven fix is rung 3 (drop prior-video ref, loose storyboard framing, non-identifiable humans). |
timed out waiting for response (600 seconds...) |
Stale build — Create Video must use the 20-minute timeout. Rebuild/redeploy oracle dist; the tv runs a dist rsynced from local, never a git build on the tv. |
| Quota or daily limit text | Stop. Report quota status and the visible message. |
| Music/dialogue appears in a clip | Do not regenerate asking for "less" music/dialogue. Treat the clip audio as unusable, strip audio in the stitch, and tighten future prompts with the Audio constraints block. |
| Character speaks / repeats a line | Same as music: characters should never talk. Strip audio in the stitch; keep the no-dialogue constraint in every prompt. |
| Clip shorter than 10s | Prompt must say "full 10 second" explicitly; re-check before regenerating. |
| Consent dialog | Click only exact Agree; never click generic first buttons. |
| Run waits forever, prompt still in composer (no user turn) | Unsubmitted send misread as accepted. Cancel locally + remotely, verify oracle includes the submit-acceptance fix, rerun only that clip. |
| Blob/fetch capture failure | Inspect the generated <video> source and use the Download video CDP fallback if implemented. |
| Remote transfer cap | Report the cap clearly; do not silently truncate or omit the video. |
Return:
ffprobe receipt paths with duration/dimensions/size;---
name: oracle-video-create
description: >
Generate videos from prompts with Gemini's Create Video / Gemini Omni tool through
Oracle's browser engine, including image-conditioned clips and multi-clip
stitchable stories with continuity references. Use when the user asks to create a
video, generate video, make a Gemini Omni clip, build a video story from images,
or use Oracle/Gemini Create Video.
---
# Oracle Create Video Pipeline
You are a PIPELINE EXECUTOR. Your job is: prepare a reproducible prompt and
optional reference images/videos -> call Oracle's Gemini Create Video browser tool ->
persist the result payload and video artifact -> verify the local MP4/WebM/MOV
with `ffprobe` -> summarize the exact paths. For multi-clip stories, you also own
the continuity chain, the stitch, and the visual sanity check.
## Proven Capabilities & Limits (verified 2026-06-10)
- **Max ~10 seconds per generated clip.** Asking for longer does not produce longer
clips. A "one minute video" means a multi-clip story: N x 10s clips + local stitch.
- Output is **1280x720, H.264 video + AAC audio** MP4 (a 10s clip is ~2 MB).
- Generation takes **2–15+ minutes**. The Create Video timeout is 20 minutes
(enforced in both the MCP tool and the CLI default); never assume failure earlier.
- `files` accepts **multiple reference images** (PNG tested) AND **previously
generated MP4s** as continuity references in the same call — Gemini uploads all of
them through the same attachment path before submit.
- Continuity chaining works: feeding clip N's MP4 + the next story images into clip
N+1 preserves character/style/grade — but identity-continuity language carries
guardrail risk (see the guardrail ladder below).
- The remote run executes on the tv host from `~/.oracle/config.json`
(`browser.remoteHost` + `browser.remoteToken`). The saved video lands LOCALLY at
`~/.oracle/sessions/{slug}/videos/video.mp4`; logs at
`~/.oracle/sessions/{slug}/output.log` and
`~/.oracle/sessions/{slug}/models/gemini-create-video.log` (there is no `log.txt`).
## Non-Negotiables
- Use browser mode only. Never use API mode for video generation.
- Use the exact MCP tool name `mcp__oracle__create_video`.
- Slugs must be 3-5 hyphenated words.
- Every prompt starts with `[{slug}]` on its own first line, then a blank line.
- The MCP return can be truncated; persist the full result JSON, run log, and video file yourself.
- Gemini Omni generation can take minutes. Do not declare failure just because a poll is slow; reattach with `mcp__oracle__sessions` before deciding.
- **Silence policy (mandatory):** generated clips must be visually driven and
stitch-safe. Characters never talk. No music, no score, no singing, no lyrics, no
dialogue, no voiceover, no spoken words, no mouth-synced speech. If Gemini insists
on audio, allow only minimal non-musical ambient/environmental sound. Even
"successful" speech ruins stitching (e.g., a character repeating the same line
twice). Every prompt MUST carry the Audio constraints block verbatim.
- Generate story clips **sequentially**, one at a time: each clip may depend on the
previous clip's MP4, and the remote browser host should hold one Create Video run
at a time.
- If Gemini shows a consent dialog, automation may click only the visible action whose normalized text is exactly `Agree`.
- If Gemini reports quota or guardrail/refusal text, stop that prompt and apply the
guardrail ladder. Do not retry-loop the same prompt.
- Close the Gemini tab only after the video file is saved and verified.
## Model / Intelligence Boundary
This skill uses Gemini Create Video through Oracle. Do not manage ChatGPT's
Intelligence pill here.
Critical rules:
- Do not set ChatGPT `browserThinkingTime` for `mcp__oracle__create_video`.
- Do not try to switch between ChatGPT `Instant` and `Pro Extended` for Gemini
video generation; those labels are ChatGPT-specific.
- The only relevant mode is Gemini Create Video / Gemini Omni as driven by the
`mcp__oracle__create_video` tool.
- If the user asks about the Instant vs Pro Extended policy while generating
video: Instant is for ChatGPT image generation; Pro Extended is for ChatGPT
non-image consults; Gemini video follows this skill's Gemini tool path.
## Single-Clip Pipeline
1. Choose a slug: 3-5 hyphenated words, e.g. `hero-loop-video-proof`.
2. Create a run folder:
```bash
RUN_DIR="runs/$(date +%Y-%m-%d)-{slug}"
mkdir -p "$RUN_DIR"/{receipts,prompts,results,videos}
```
3. Write the prompt to `$RUN_DIR/prompts/create-video.md`:
```text
[{slug}]
{video generation prompt — include "full 10 second" explicitly for max length}
Audio constraints: no music, no score, no singing, no lyrics, no dialogue, no voiceover, no spoken words, no mouth-synced speech. Prefer a silent clip; if audio is unavoidable, use only subtle non-musical ambient sound.
```
Also end visual prompts with: `No text, no captions, no readable words.`
4. Run one video generation:
```
mcp__oracle__create_video({
prompt: "[{slug}]\n\n{video generation prompt}\n\nAudio constraints: no music, no score, no singing, no lyrics, no dialogue, no voiceover, no spoken words, no mouth-synced speech. Prefer a silent clip; if audio is unavoidable, use only subtle non-musical ambient sound.",
slug: "{slug}",
files: []
})
```
For image-conditioned video, pass absolute paths (order matters — describe them as
"sequential visual references" and reference them in order in the prompt):
```
mcp__oracle__create_video({
prompt: "[{slug}]\n\nUse the attached images as sequential visual references. {prompt}\n\nAudio constraints: ...",
slug: "{slug}",
files: ["/absolute/path/reference-01.png", "/absolute/path/reference-02.png"]
})
```
For continuity from a prior clip, put the previous MP4 FIRST in `files` and say
"Use the attached previous video as continuity reference for the camera language,
color grade, and tone" (see guardrail ladder before using identity language).
Optional debugging only: `browserKeepBrowser: true`, `browserConversationUrl: "..."`.
CLI recovery shape (same semantics; useful for batches and when MCP is unavailable):
```bash
oracle --engine browser --model gemini-create-video --slug {slug} \
--prompt "$(cat "$RUN_DIR/prompts/create-video.md")" \
--file /abs/ref-01.png --file /abs/ref-02.png \
--verbose --wait --force
```
5. Persist artifacts:
- `$RUN_DIR/results/create-video-result.json`
- `$RUN_DIR/results/create-video-output.log` (tail of `~/.oracle/sessions/{slug}/output.log`)
- copy every returned `videoPaths[]` entry (typically
`~/.oracle/sessions/{slug}/videos/video.mp4`) into `$RUN_DIR/videos/` with a
descriptive name.
6. Verify the video:
```bash
ffprobe -v error -show_entries format=duration,size:stream=codec_type,codec_name,width,height,duration \
-of json "$RUN_DIR/videos/{name}.mp4" > "$RUN_DIR/receipts/ffprobe.{name}.json"
```
Duration must be greater than zero and should match what was requested (e.g.
`10.000000` for a full-length clip). Include codec, dimensions, duration, and size
in the final summary.
Healthy run log markers: `Gemini Create video tool selected.`,
`Attachment upload complete in Gemini: <each file>`, `files=N`,
`Gemini video still generating... (generating)`, `Saved 1 generated video to ...`.
## Multi-Clip Story Pipeline (stitchable ~1 minute story)
Proven shape: 10 story images -> four 10s clips -> 40s straight concat -> 60s paced edit.
1. One run folder for the whole story:
`runs/YYYY-MM-DD-{story-slug}/{prompts,results,receipts,videos}` plus a
`manifest.md` mapping each segment to its images, continuity reference, prompt
file, and Oracle session slug. Write ALL per-clip prompt files up front.
2. Group source images into beats of 2-3 images per 10s clip. Each clip's prompt:
"Use the attached previous video as continuity reference... Use the N attached
images as the next sequential story beats. Generate a full 10 second cinematic
video: {beat-by-beat description}." + the Audio constraints block + "No text, no
captions, no readable words."
3. Generate sequentially. After each clip: copy
`~/.oracle/sessions/{clip-slug}/videos/video.mp4` ->
`$RUN_DIR/videos/NN-{beat-name}.mp4`, save the output.log tail, write the ffprobe
receipt, and only then start the next clip with this MP4 as its first `files` entry.
4. Stitch (clips share codec/resolution, so stream-copy concat works):
```bash
cd "$RUN_DIR"
cat > concat.txt <<'EOF'
file 'videos/01-....mp4'
file 'videos/02-....mp4'
...
EOF
ffmpeg -hide_banner -y -f concat -safe 0 -i concat.txt -c copy {story}-40s.mp4
# Optional ~60s paced edit (1.5x duration; re-encode):
ffmpeg -hide_banner -y -i {story}-40s.mp4 \
-filter_complex '[0:v]setpts=1.5*PTS[v];[0:a]atempo=0.6666667[a]' \
-map '[v]' -map '[a]' {story}-60s.mp4
```
For final deliverables, strip generated audio in the stitch (`-an`) unless the
user explicitly asks to preserve it; add music/narration/sound design only after
the picture edit is locked, as one continuous pass.
5. Visual sanity receipt — contact sheet, then actually look at it:
```bash
ffmpeg -hide_banner -y -i {story}-60s.mp4 -vf 'fps=1/5,scale=320:-1,tile=3x4' \
-frames:v 1 receipts/contact-sheet-60s.jpg
```
6. Write `SUMMARY.md` in the run folder: outputs, per-clip sessions/images/refs,
ffprobe receipts, and any guardrail notes.
## Guardrail Ladder: "specific people" refusals
Gemini may refuse with text like **"I can't make content that appears to show
specific people."** This fires probabilistically on identity-continuity prompts —
the same shape (prior MP4 + person images + "same man") can succeed for several
clips and then fail. When it fires, do NOT retry the same prompt. Step down:
1. **Full continuity** (default first attempt): prior clip MP4 + next images +
"preserve the same protagonist / same character design".
2. **Fictional reframe**: keep attachments, describe everything as "a fictional
short film", references "for wardrobe, framing, lighting, color grade, and tone",
"without identifying or recreating any real person". (This alone has FAILED
against the specific-people guardrail — don't expect it to be enough.)
3. **Proven fallback**: DROP the prior-video attachment entirely. Use only the
images "as loose storyboard references for scene layout and emotional tone, not
as identity references", and require non-identifiable human framing: "If a human
appears, show them only from behind, in silhouette, at a distance, or with the
face obscured and not identifiable." Let objects/setting/grade carry continuity.
Record which rung succeeded in the run notes so the stitch summary explains any
continuity seams.
## Wait, Inspect, Cancel
- Healthy waiting shows `(generating)` polls. **`(complete-without-video)` polls
mean Gemini answered with text but produced no `<video>` — almost always a
guardrail/refusal.** Do not wait out the 20-minute cap: inspect the live run
target, read the visible response text, then cancel and apply the ladder.
- Inspecting the live run target on the tv: `scripts/browser-tools.ts eval` and
`/json/activate` only reach the *active* tab, which is often NOT the isolated run
target. Instead: `curl http://127.0.0.1:<chrome-debug-port>/json/list` on the tv,
identify the Gemini page target whose body contains this run's prompt text, then
connect straight to its `webSocketDebuggerUrl` with Node's built-in WebSocket
(Node >= 22; no `ws` package needed) and send `Runtime.evaluate` to read
`user-query` / `model-response` counts and visible text. A genuinely submitted run
shows 1 user-query, 1 model-response, and text like "I'm generating your video.
This could take a few minutes".
- A run that waits forever with the prompt still sitting in the composer (no
user-query turn) is an UNSUBMITTED send that the wrapper mistook for accepted —
this was a real bug (fixed in oracle `f1d6231b` by requiring a real submitted
turn/response or prompt disappearance). If you see it again, treat it as a send
failure: cancel, fix/report, rerun.
- Cancelling a stuck/guarded run (both sides):
1. Kill the local CLI process (`ps -axo pid,command | rg 'gemini-create-video'`).
2. Cancel remotely: `curl -X POST -H "Authorization: Bearer $TOKEN"
http://{remoteHost}/runs/{runId}/cancel` (find the runId via `GET /runs`).
- **Idle gate before starting a run: use `GET /runs` and check `.active` is empty.**
The `/health` `activeRuns` counter can leak after cancellations and stay nonzero
while no run is actually active.
## Troubleshooting
| Symptom | Action |
| --- | --- |
| `create-video-tool-not-found` | Gemini menu DOM changed. Re-probe the remote Gemini tab and update selectors before retrying. |
| `create-video-chip-missing` | The menu click did not activate the `Videos` composer chip. Retry once after inspecting the composer DOM. |
| Polls show `(complete-without-video)` | Guardrail/refusal in the response text. Inspect the run target via CDP, read the message, cancel both sides, apply the guardrail ladder. |
| "can't make content that appears to show specific people" | Apply the guardrail ladder; the proven fix is rung 3 (drop prior-video ref, loose storyboard framing, non-identifiable humans). |
| `timed out waiting for response (600 seconds...)` | Stale build — Create Video must use the 20-minute timeout. Rebuild/redeploy oracle dist; the tv runs a dist rsynced from local, never a git build on the tv. |
| Quota or daily limit text | Stop. Report quota status and the visible message. |
| Music/dialogue appears in a clip | Do not regenerate asking for "less" music/dialogue. Treat the clip audio as unusable, strip audio in the stitch, and tighten future prompts with the Audio constraints block. |
| Character speaks / repeats a line | Same as music: characters should never talk. Strip audio in the stitch; keep the no-dialogue constraint in every prompt. |
| Clip shorter than 10s | Prompt must say "full 10 second" explicitly; re-check before regenerating. |
| Consent dialog | Click only exact `Agree`; never click generic first buttons. |
| Run waits forever, prompt still in composer (no user turn) | Unsubmitted send misread as accepted. Cancel locally + remotely, verify oracle includes the submit-acceptance fix, rerun only that clip. |
| Blob/fetch capture failure | Inspect the generated `<video>` source and use the `Download video` CDP fallback if implemented. |
| Remote transfer cap | Report the cap clearly; do not silently truncate or omit the video. |
## Final Summary Shape
Return:
- slug(s) and run directory;
- Oracle session id(s);
- video path(s) — per-clip and stitched outputs;
- `ffprobe` receipt paths with duration/dimensions/size;
- contact-sheet path for multi-clip stories;
- which guardrail-ladder rung each clip needed, and any quota, guardrail, or DOM
facts that future runs need.