How to use
Pinacle Studio.

CHAPTER 01Start here

What Pinacle Studio is, what it needs from you, and how to get around it.

Pinacle Studio turns a topic into a finished YouTube video. It finds something worth making, writes the script, arranges your voiceover, plans an image for every couple of seconds of narration, draws those images in Google Flow, places them on the audio and renders the file - then writes the description, the tags and a thumbnail to go up with it.

It is a Chrome extension and everything runs on your machine. There is no AI service of ours in the middle: the studio drives ChatGPT, Claude, Gemini and Google Flow inside your own signed-in browser sessions, and the pictures come back to your own Flow canvas.

The rule it holds itself to

Nothing is claimed before it exists. A timestamp is never invented - where the studio cannot measure one, it says the timing was estimated. A preview is never upscaled to look like a render the model did not make. An image is reported finished only once its bytes have been read back. When something fails, the Activity card says what failed and what it cost, in words rather than a shrug.

What you need before the first project

Chrome with the extension loaded, and a Google account signed in - the Images stage draws with Flow in your own session.

Two to five channels in your niche for the Ideation stage to read. Your own channel is optional, but the Publish stage reads its past descriptions when you have it.

The voiceover audio. The studio does not speak: record it yourself, or export it from any text-to-speech tool you like, then bring the files in. One file per section is the normal way.

A free Groq API key if you want a transcript in seconds rather than minutes. Without one, transcription runs on this machine with Whisper and takes as long as it takes.

One project, one video

Everything - channels, script, audio, prompts, pictures, the timeline - is stored per project in this browser. Settings, Projects renames the project you are in and creates or switches to another. Your data, in the same place, exports a project as a single .zip and imports it back; that is how a project moves between machines or survives a reinstall.

A project lives in the browser profile that made it, pictures and all. To open one somewhere else, export it from here and import it there.

How to get around

The rail on the left is the order of work: Ideation, Script, Voiceover, Images, Timeline, Publish. How to is this page, and Settings is the bottom of the rail.

You can explore each section before its inputs are ready. The Timeline opens even when empty; add a combined voiceover and image plan before building and exporting. Some workflow handoffs explain any unfinished prerequisites.

The step strip at the top of each stage shows what is done (a tick), what is active now, and what is still to come.

Progress bars and status lines are real: which picture is being drawn, which half of a research run is running, how much is left.

The Activity card at the bottom of every stage is the log. When something looks stuck, that is the first thing to read.

CHAPTER 02The six stages at a glance

The whole pipeline on one table. Each stage leaves behind what the next one reads.

Stage What it does What it leaves behind

Ideation Reads the channels you name and writes titles from what is already working A title

Script Deep research, a style sample and the videos already winning the topic, then a script of a set length The script and its target duration

Voiceover Orders your audio files against the script, combines them into one file and transcribes it One audio file and a timestamped transcript

Images Plans a prompt per couple of seconds of narration, draws them in Flow and collects the files back Prompts, and pictures placed on the timeline

Timeline Puts the pictures on the voiceover, checks the result, renders the video The finished video file

Publish Writes the description, the hashtags and the tags that file the video, and draws a thumbnail from a frame the video already has The words that go up with the video

Time is the axis through all of it. The voiceover is the clock, every image is placed on a window of it, and the Timeline is that same clock with pictures on it - which is why the stages have to be done in this order.

Reset appears near the top of each stage and empties that stage: the Ideation channel list and scrape, the Script setup and draft, the imported voiceover, the Images plan and pictures, the built timeline and edits, or the Publish package. Each asks first. Settings' Reset everything clears the whole project.

CHAPTER 031. Ideation - find something worth making

Read what is already working in your niche, then have titles written from it.

Add your competitors

Press Add under Competitors and paste a channel URL - youtube.com/@channel is enough. You can use 1 to 20 channels. The scan reads each channel's own page, and the profile card under the rail is your own channel.

Say how much to read

Videos per channel is how many of each channel's recent videos are scraped. Titles is how many titles the model writes. Sort decides whether the scrape takes the most popular or the most recent.

Describe what you want

Prompt holds the title-writing instructions. With no channel or video references, the input says Your rough idea. Once you add references, it becomes Anything for the model to know and holds your direction for those sources.

Ideate from videos

Open the card, enter a working title and search YouTube. Press Ideate on up to 20 video cards. Return here to review their thumbnails and titles. The selected videos can be combined with your competitor channels in the same title run.

Press Generate titles

It fetches every selected channel, combines that data with any selected videos and your direction, then asks the model for titles. With no channels, it uses the videos or idea you supplied. Cancel stops a running job.

Read what came back

Scraped data is what was really found on each channel. Titles is what was written from it. Regenerate asks for a fresh set, Add more asks for more on top of the ones on screen, and Copy all puts the list on your clipboard.

Send one to the Script stage

Use the pencil on a title row to edit its wording, then Save. Script carries the saved title into the next stage.

Titles use the sources you provided: fetched channels, selected videos and your own direction. If a channel URL was wrong or a page had changed shape, check Scraped data before generating again.

The cards fold. Each one opens by itself while it holds something - channels on Competitors, text in the rough idea or direction input, chosen videos on Ideate from videos, your own words in the Prompt or the Script style reference - and stays folded while it is empty. Press a card's heading to fold it yourself; your choice is kept.

CHAPTER 042. Script - four inputs, one script

Decide the length, gather the facts, set the voice, and let the model write.

Script setup holds the three decisions that shape the draft: Title (type one, or bring one over from Ideation), Target duration (minutes and seconds) and Speaking speed (words per minute - 150 is a calm narrator, 180 is brisk). After a competitor scrape, Target duration uses the time shown on the most-viewed scraped video card, rounded up to the next whole minute. You can edit the minutes and seconds; your choice stays in place. Scrapes made before this feature need to be run again to collect those card times. The word count the script has to hit is worked out from those three and shown before you generate. Press the Script setup or Style reference heading to fold it; the fold stays where you leave it.

Step 1 - Deep research

Press Deep research and the studio has ChatGPT's Deep Research and Gemini research the topic, each in its own tab. Copy prompt and Open ChatGPT / Open Gemini are the manual door if you would rather drive them yourself, and Paste research takes anything you already have. The estimate beside the heading - about twenty minutes - is honest: this is the slow step, and the studio now waits up to forty-five minutes for a pass before it calls it lost.

Step 2 - Style reference

Paste a script, a transcript or any text whose speaking style the new script should learn. It is optional, and it changes how the writing sounds, never what it is about.

Step 3 - Similar videos

The videos already doing well on this topic, so the script is written knowing what it is up against.

Press Generate script

The model writes the draft to the word count. Progress follows the run and Cancel stops it; whatever came back is kept.

What you get back

The Script card holds the draft with a live word count against the target.

Copy puts the whole script on the clipboard.

Script Master Prompt starts empty for optional instructions. The built-in rule studies your references' structure, openings, pacing and endings, then combines their strongest techniques into an original script for your video. It never copies wording or jokes. Reset clears your additions and keeps that default rule.

The script as written is kept - but it is NOT what the Images stage plans from. That stage plans from the narration as spoken, read off the voiceover, so the pictures follow the audio rather than a draft you have edited since.

CHAPTER 053. Voiceover - the audio becomes the clock

Bring the audio in, put it in order, combine it, and read it back with timestamps.

Bring the audio in

Drop your files on the drop zone, or press Choose files. Record the script yourself, or export it from a text-to-speech tool. One file per section is the usual way.

Say what is being said

Paste the script into The script this narrates, or import a .txt. Every segment is compared against the line it should be reading, which is how the stage can tell you when a file and a line disagree.

Arrange it

Auto arrange by the script orders the files to match the script's own order, so you do not have to name them in order. The Review table shows every segment against its line; one row can be transcribed on its own, and Compare shows the differences between what a segment says and what the script says.

Combine and listen

Press Combine the segments. It joins the files into one voiceover, cross-fading each joint (six milliseconds by default) so the seams do not click. The advanced panel holds the engine: Local Whisper works on this machine, Groq cloud is much faster and needs the free key from Settings. Download the merged file keeps a copy for yourself - the studio keeps its own.

Leave trimming on

The trim switch under the combined file removes the dead air at the start and end of each clip. Leave it on: silence left at a clip's edges is what puts pictures slightly ahead of the words on the Timeline later.

Transcribe it

Press Transcribe in the Timestamped transcript card. This read is what every later stage works from. Download SRT or Download JSON if you want it elsewhere, or Copy it as text.

Continue to Images

The button at the bottom. It refuses until the combined file has passed its check and the transcript exists - which is why it is worth pressing rather than fighting.

Measured or estimated

Measured - transcribed by Whisper or Groq, so every mark is a position the audio really has: the words themselves were timed.

Estimated - derived without word timings (a pasted transcript, for example). Still usable, and every stage downstream says so instead of pretending otherwise.

One image per two seconds of narration is the planning target the Images stage suggests, not a rule. The narration decides how many pictures the video needs; the target only frames the suggestion.

CHAPTER 064. Images - a prompt per frame, drawn and collected

Plan the visuals against the voiceover, draw them in Flow, and file them against the prompts that asked for them.

Subject - who or what must stay the same

The switch on the Subject card turns the whole consistency workflow on. With it off, every recurring person has to be described in full in every prompt instead.

When the subject appears: Every scene pins your reference pictures to every prompt; Selected attaches them where the narration is about the subject; Only if told attaches them only where the model marks a line with @subject.

Subject references: up to four pictures. They are what the subject IS - there is no description to write.

Create the prompts

Pick the master prompt

Two documents ship: the flat 2D doodle stickman explainer and the hand-inked ancient-Rome storybook explainer. The switch at the top of this card chooses which one the whole project is written with - it is remembered with the project, and it decides what the box below opens with.

Style is the master prompt

The chosen master prompt is the look every written line has to carry. Open the box, edit it for a different look, and press Reset to put the shipped one back.

Check the narration

The Script card in this stage fetches the timestamped transcript from the Voiceover stage, and Preview shows exactly what will be handed to the model. If the transcript or the script changes afterwards, fetch it again: this stage will not plan from a stale read.

Choose the batch size

Prompts per batch is how many prompts ONE press asks for: up to 20 on trial or 100 on Pro. Choosing 60 requests 60 image prompts without forcing them to cover the entire supplied narration. The model chooses their timings and holds, and a reply with a different count is accepted. The next batch continues from its last image end.

Write them

Press Generate image prompts. The master prompt, the narration for this stretch and your settings go to the model chosen in Settings. Copy the prompt and Paste the reply are the manual door: copy it into any model you like, paste the reply back, and press Read reply.

Read the queue

The Prompt queue lists what came back, newest sitting first. Each row holds the prompt, the narration under it, its own draw controls and Replace.

A long video is written in batches

A batch is N prompts from where the last one stopped. Pressing again APPENDS the next stretch - the button says which mark it starts at, like "Generate the next batch (20 from 00:01:00)".

The count box comes down to the beats the narration has left once a reply lands, and says so, so the next press asks for what can actually be granted.

Ids are minted past every id already in use, so a reply that numbers its images from 1 again cannot make a second image_001. Nothing is drawn twice.

The last ten prompts of the plan ride with every continuation batch. They are the record of how each recurring person looks: the block tells the model to COPY a repeated person's description into the new line in the same words, instead of writing "the same as before".

Recurring people can be named once and called back: the model defines them with #Name and calls them in a line with @Name. Each is drawn once and attached only to the scenes that call it. With the Subject switch off none of that runs, and every person is described in full every time.

Regenerate batch clears the newest sitting and asks for the same stretch again. Start over rewinds the whole plan to zero and stops a run in flight. Reset, at the top of the stage, is the one that wipes it - prompts, pictures, subject, characters - and it asks first.

Draw them in Flow

Generate the missing images hands the prompts to Google Flow in narration order, one at a time, and files each finished picture into Collect as it lands. Stop after this one finishes the image in flight, then stops.

Check reads Flow's own canvas - model, aspect, images per prompt, quality. The studio only ever READS these; set them by hand in Flow. Two of them cost real renders: Images per prompt should be x1, and the aspect must be 16:9.

Per prompt you can Start this one, Skip it, Replace it with a file of your own, or Clear image - which forgets the picture and never touches the prompt.

Flow opens in a background tab by default and is never brought forward, except once in a run if the canvas has stalled.

Collect the pictures

Drop your whole folder of renders on Collect images, or press Choose. Nothing has to be renamed by hand.

Match & rename matches files to prompts by the names inside the folder and shows the plan: which file goes on which prompt, and which files belong to no prompt. Save renamed copies writes the renamed set out.

The names are placement and hold: 003_00:12-00:15_3.0s is the third frame on the timeline, holding from 12 to 15 seconds, for three seconds.

A picture the studio cannot match is reported, never guessed at - and no picture is claimed by name alone. A file belongs to a prompt only when the plan says it is that prompt's.

CHAPTER 075. Timeline - pictures on the clock, and the render

The voiceover is the clock and the pictures sit on it. When the two agree, render.

The bar at the top says how far along the video is: how many pictures are drawn out of the frames placed, the seconds they cover out of the voiceover, and the percentage that comes to. The meter under it is the same read as a shape.

The preview shows the frame at the playhead with the narration spoken over it. Use its top-right fullscreen button for a large view and Escape to return. The transport plays it: previous and next frame, play, speed, volume.

The track is the whole video - the picture lane and the voiceover wave - with one playhead across both. Drag the playhead, or press a frame to inspect it.

The frame inspector shows the prompt a frame was drawn from and lets you move its start and end, move it earlier or later, take it off the timeline, or Delete the picture everywhere, bytes and all.

Moves, trims and extensions update the preview and the next export. Empty gaps play black until filled. If clips overlap, the most recently edited clip appears over the overlap; the checks identify it.

Effects and extra audio

Import audio places music and sound files at the playhead. Audio clips can be moved or trimmed with their edge handles. Select a clip for precise position, source in/out, speed, volume, mute and removal controls. Audio files and edits are saved with the project.

Effects, Visual transitions contains Zoom in, Zoom out and mixed zooms. Zoom amount controls magnification. Animation duration is the percentage of the image window used for the movement.

Effects, Audio offers Pop and Whoosh. Count defaults to five. Yes randomizes placement; No spreads sounds evenly. Apply sounds replaces the generated set. Remove effects clears that set without removing imported audio.

Video speed changes the entire composition. Narration volume and each audio clip’s speed, volume, mute and trim settings also apply to the export.

Checks - read them before you render

Gaps, overlaps, frames with no picture file, and a length that disagrees with the audio are all reported, in seconds, with the frame named.

The one that catches most people: pictures arriving slightly early. When most frames begin before the first mark inside their window, the check says so and names the cause - edge silence left in the merged voiceover. Combine the voiceover again with trimming on, or move every picture at once.

Nothing is fixed behind your back. Problems are shown and left standing so you can decide.

Export and render

Choose Resolution, Frame rate, Quality and Format. Estimated size is a rough number, not a promise.

Press Render video. Progress names the stage it is on and the percentage, and Cancel stops it.

When it finishes the video plays in the panel - read back from the bytes, not assumed - its real length is measured, and Save the video writes it out. Further edits mark that video outdated; render again before saving.

Video speed re-times pictures, narration and extra audio together. Review the preview after changing it, then render the updated composition.

The render is the one place a mistake is expensive, so the stage refuses to claim success early: a file is reported only after it has been loaded and measured.

CHAPTER 086. Publish - the words that go up with the video

A description, the hashtags, the tags that file it and a thumbnail, all written from the video this project already has.

Read your channels

Descriptions from your channels reads the descriptions of your own past videos, and Most popular videos per channel decides how many of each to read. Add a channel takes any channel URL. This is what the writing is written FROM: what already works for your audience.

Videos you add by hand

Add thumbnail sits at the end of a video row - here, in Ideation, or on the video cards on YouTube itself. It puts that video in the Thumbnail card as a reference and queues its description for the next read. Added videos are read alongside the channels and are never counted with them.

Write the package

Press the write button on The package and the model answers with the description, the hashtags and the tags in one reply. Paste the reply is the manual door, as it is in every other stage.

Read and copy

Description, Hashtags and Tags each show what was written, with their own Copy button. Hashtags are the #words a viewer sees and YouTube reads only the first few, which the card counts. Tags are the plain keywords YouTube files the video under, written as one comma-separated line because that is what Studio's Tags box takes - Copy gives you exactly that line.

Draw the thumbnail

The Thumbnail card writes concepts, then draws each one in Flow at a repeatable size. Words on the picture, a headline and any extra direction are yours to set first. References are frames the video already has, and Add your own takes a picture of yours.

Pick one and see what goes up

Choose the concept you want. What goes up lists everything - description, tags, thumbnail - with Copy everything and Download JSON. The last button is the only one that opens YouTube Studio; nothing is uploaded from the studio.

Nothing in this stage touches the video. The video was finished in the Timeline; this is only the words around it. Reset, at the top of the stage, deletes everything in it - the package, the pictures it drew, the descriptions that were read and the videos you added by hand - and nothing outside it.

CHAPTER 09Settings, projects and your data

The switches that change how the stages behave, the one key the studio can use, and where the work is kept.

Dark mode changes every studio section. Turn it off for a light interface. The preference is saved across visits and shared across projects.

Hide error messages suppresses failure alerts. It is a destructive option because failed actions and unsaved work may go unnoticed. Turn it off to show new alerts again. Validation checks and job progress remain available.

First-visit spotlight tours explain each section. Skip onboarding stops automatic tours. Use Play onboarding here in How to to return to any section and replay its tour.

Skip confirmation popups: with it on, no "are you sure?" is drawn anywhere - including the ones that delete. Every skipped answer still goes to the Activity log.

Which model answers what: titles, research, scripts, image prompts and publish each pick their own model (ChatGPT, Claude or Gemini), so a stage can use the one that is best at it. Titles is one row for both ways of getting them - the competitor run and Ideate from videos - and research is one row for both passes it runs.

The Flow canvas: Open Flow in decides whether Flow works in a quiet background tab or a small window, Wait between prompts paces a long run, and Extra tries per image is how often a failed render is tried again before it is reported.

Save folder, inside Downloads and Image name style decide where finished images are written and what they are called, when saving to disk is on.

Groq API key is the one credential the studio uses. It is free, it is stored in this browser, and it turns transcription from minutes into seconds.

Projects renames the project you are in, and creates or switches to another. One project per video is the intention.

Your data exports a project as one .zip - every stage, the images, the audio, the settings - and imports it back. Importing replaces what is in this browser, which is why it asks.

A project lives in the browser profile that made it. Export it whenever you want it somewhere else.

CHAPTER 10When something goes wrong

The symptoms that come up most, what they mean, and what to do about them. The Activity card always has the longer version.

The model replied but the batch was not read

Use Paste the reply: paste the reply into the box on the Images card and press Read reply. A reply that is not in one of the three shapes the reader knows - windowed lines, scene blocks, or one JSON object - is reported rather than half-read.

Flow never enabled its send button

The words reached the page but Flow did not accept them, so nothing was sent and nothing was spent. Try again, and check that Flow is signed in with a project open.

The images come back the wrong shape

Flow's aspect must be 16:9. The studio reads it and warns about it, and never presses the control itself. Images per prompt should be x1 as well, or every collected frame cost four renders.

The pictures come in a moment before the words

Edge silence in the merged voiceover. Combine the voiceover again with trimming on, or use the shift control on the Timeline to move every picture at once. The Checks name the fault in seconds.

A batch came back short

The count box comes down to the beats the narration has left, and says so. Regenerate batch asks for the count the sitting was originally asked for, so a short batch can still be asked for again in full.

Choose a new chat for the next batch

Batches normally reopen the same conversation for this project. Press New chat for next batch, then Generate, to use a fresh conversation while keeping your prompts, images, timing and timeline. Carry the last 10 prompts only controls the text quoted in the next ask; it does not choose the chat.

The Timeline says it needs Voiceover and Images first

It needs a combined voiceover (stage 3, second step) and at least one prompt in the Images plan. The note in the corner names whichever is missing.

A frame says it has no image file

The picture lives in this browser's storage. If the project was imported on another machine, or browser data was cleared, the bytes are gone - the prompt is untouched and the frame simply needs drawing again.

Something is stuck and nothing says why

Read the Activity card. It is the log of every job and the receipt for anything skipped or failed.

CHAPTER 11The words the studio uses

Short definitions, so the cards are readable.

Batch

A count of image prompts from where the last batch stopped - not the number of images in the video.

Beat

A short stretch of narration, up to a few seconds, that one idea is spoken over. The narration is cut into beats, and every image covers one or more of them.

Cursor

The mark in the narration where the next batch starts. It follows the reply that came back, not the request that was made.

Hold

How long one picture stays on screen. The model decides it, guided by "shorter is better": the picture changes when the narration moves.

Placement name

The filename a picture is given from where it sits, like 003_00:12-00:15_3.0s. That is what lets a folder of renders be matched back to the prompts that asked for them.

Sitting

One press of the generate button and the batch it produced. The queue is read in sittings, newest first.

Master prompt

The document a model writes from: the script prompt, the image batch prompt (which is also the style) and the publish prompt. All three are editable.

Style lock

The part of an image prompt that carries the look. The same rules in every line, so the whole video looks like one video.

Subject reference

Pictures of a person or thing that must stay the same, attached to the prompts that should have them.

#Name and @Name

How a recurring person is named once and then called back, so each is drawn once and attached only to the lines that call it.

Transcript

The timestamped narration the Images stage plans from. Measured when it was transcribed from the audio, estimated when it was derived without word timings.

Ledger

The record of what has been drawn: one entry per prompt, with its state, how many tries it has had, and where its picture lives.