Bloom Studio
Video Editing Tools for Communities
Put a recording in your inbox folder and type /process to your AI. It writes one page showing who spoke, when, and which lines are worth cutting. From there, ask it for stills, clips or a posting plan.
- Runs on your computer
- Nothing uploaded or posted
- You approve every cut
What to ask for
Five things to ask for.
Type one of these to your AI while it has the Bloom Studio folder open.
/processReads what is new in your inbox and ends on the readthrough page.
/planA posting plan for one recording: which clip goes out on which day, with one ask and a caption for each.
/makeStills, a clip, a tightened take or captions, from any recording. It shows you a plan and one frame before it renders.
/updatesChecks for newer tools and gets them when you say yes. It waits for any job still running.
/helpWhen something goes wrong: a short report to send to Megan. She replies with a prompt you paste back.
End card and QRComing soonA closing card with a scannable link, added to a finished clip.
How it works
From a recording to a posting plan.
Each step saves a file the next step reads, and stops wherever you have a decision to make.
Stage 1 · You
You add a recording
Put the video in your inbox folder, with its captions if you have them.
- your turn
Type /process
Stage 2
Your AI reads it
It makes any missing captions, checks the sound, reads the transcript and pulls a still per section.
Makes readthrough.html
Stage 3 · You
You read one page
You correct the names on it and pick the lines worth cutting.
- your turn
Makes a prompt for your AI
Stage 4
You ask for what you need
Ask for stills, a Reel, a shorter take or captions. You see a plan before anything is made.
Makes clips and frames
Start here
Get running on your computer.
You do this once. You get the folder and an AI assistant, and the AI installs the rest. Windows only for now.
Phase 1 · You
Get the folder and an AI
- Join the pilot on GitHub
- Clone it with GitHub Desktop
- Install Claude, open the folder
Phase 2 · Your AI
Sets it up from one prompt
- Installs Python, FFmpeg and Node
- Makes Video Work and its inbox
- Points the start page at it
Phase 3 · You, then your AI
Your first recording
- Drop it in the inbox
- Type /process
- Open the readthrough
Private while it is a pilot. Once you are added, clone it:
git clone https://github.com/HoustonBloom/bloom-studio.git
Step by step
The four steps
- You
Get Bloom Studio
Ask in the Discord to join the pilot. Once you are added, clone HoustonBloom/bloom-studio with GitHub Desktop. A clone updates itself with /updates. A zip download cannot.
- You
Get your AI assistant
Install Claude and sign in. Use Claude Code or Cowork, so it can work with files on your computer, and give it the Bloom Studio folder. Download Claude.
Show the prompt
I am setting up Bloom Studio, a set of video tools in this folder. Read AGENTS.md here, then tell me in two lines what you can do with it.
- You paste, your AI works
Set it up with one prompt
Your AI installs the tools, makes your Video Work folder with its inbox, and points the start page at it. It took 1 min 42 s and 2 min 10 s on two test runs, and checks for 950 MB free first.
Show the prompt
Set up Bloom Studio for me. In this Bloom Studio folder: run install.bat -NoPause and wait until every tool says ready. Then make a folder called Video Work in my Documents folder, with a folder called inbox inside it. Then run point.bat -Workspace "<the full path of Video Work>" -NoShortcut -NoOpen. When it is done, tell me to reopen this page.
- You paste, your AI works
Your first recording
Put a recording in
Video Work/inbox, with its captions if you have them, then paste this. With captions, plan for about 30 minutes on a 95-minute recording. Without, add about 30 more.Show the prompt
/process the new recording in my inbox
If something fails, type /help.
It writes a short report to send to Megan, and she replies with a prompt you paste back. Your AI is told to stop and name the step, not guess.
The start page
A start page that holds no data.
START-HERE.html opens from your computer and reads your Video Work folder each time. It stores nothing itself.
Setup
One slide per step, each with its prompt and a Copy button. A step ticks only when the page can see it is done.
Gallery
One poster per recording, linking to its readthrough and its posting plan.
Explore
A map of your recordings and the pages made from them.
Inbox banner
What the last check found in your inbox, with the /process prompt ready to copy.
It is built from a template in the Community Legibility Spec. You point it at your own folder, change it, and share your changes back. Nothing leaves your computer.
What you get
What a run leaves behind.
Each run saves plain files in your Video Work folder. It sits outside Bloom Studio, so updates never touch it.
- inbox/where you drop a recording
- relational-space/what the start page reads
- my-recording/one folder per recording
- readthrough.htmlthe page to read
- postplan/lineup.htmlthe posting plan, if you said yes
- stills/frames and a contact sheet
- clips/what you keep
- _agent/your AI's files: transcript, segments, logs
readthrough.html
Who spoke, when, what each part covered, and the lines worth cutting. The stills sit inside the file.
lineup.html
Posts week by week, each with a date, one ask, a caption and a status. You accept or reject each one.
contact-sheet.html
Frames from the moments where something happened, cropped for a feed. You pick by looking.
a plan, then a render
Names every segment and why it is there. Nothing renders until you approve the plan.
Measured on a 72-minute episode
- stretches
- 5
- moments
- 12
- lines proposed
- 43
Measured on a 3-hour gathering
- stretches
- 11
- moments
- 26
- lines proposed
- 93
The tools
Every tool, one job each.
Each tool works on any video by itself, and Bloom Studio includes all of them. A linked tool opens its README on GitHub.
10 of these have no GitHub folder yet, so they carry no link.
Read
Know what you have
- transcribe ToolA caption file from a recording that arrived with none.InA video fileOutA .vttOpen on GitHub
- noise-check ToolHow clean the sound is, and where it is not.InA recording or a render, optionally its planOutNumbers, a verdict and reasons, or noise.json for the readthrough page
- readthrough WorkflowRead a whole recording and say what is in it: who was up, when, and where the demo was.InCaptions, and the AI's read of the whole transcriptOutAn HTML pageOpen on GitHub
Make
Cut, frame and pick
- shorts ToolTalk footage in, a vertical Short out.InA source video and a planOutA rendered 1080x1920 clip and its receiptOpen on GitHub
- tighten ToolShorten a take without changing what it says.InA clipOutThe same clip, shorter, and its receiptOpen on GitHub
- stills ToolPhotographs out of a recording.InA videoOutImages, and a page to pick them onOpen on GitHub
- framing ToolWhere the crop and zoom go when a wide recording becomes a vertical, and a check before rendering.InA plan and, optionally, where each person sitsOutErrors and warnings about the crops
- ffmpeg-recipes ToolThe ffmpeg jobs that come up on every edit, as commands that print one line.InA video and, for the sheet, a planOutA cut, a stacked vertical, a video with a card, or one contact sheet
Deliver
Finish, hand over, plan
- loudness ToolGet a render to the loudness it should be.InAny video or audio fileOutThe same file, on target, and its receiptOpen on GitHub
- noise-fix ToolRepairs hum, hiss, rumble, clicks and mild clipping, on a copy.InA file and the findings from noise-checkOutA new file, and the measured before and after, and a receipt
- captions ToolBurned-in captions on a finished vertical video, after a person approves the words.InA finished video, optionally its plan and a names listOutA captions JSON to correct, then the captioned video, its .srt caption file and a receipt
- delivery ToolFrom a finished edit to files ready to hand over.InA finished vertical and a template pictureOutThe same video with a call-to-action end card and a QR that has been decoded, and a receipt naming the card
- postplan ToolWhen each finished video posts, what it asks, and how the caption reads.InA plan JSON, the run's clip map, the finished videos, and a calendar, and the scheduled and posted dates the owner givesOutA posting sheet, a plan page, and the lineup: postplan/lineup.html in the run folder
- clipmap ToolSay what is in a finished clip without transcribing it again.InThe plans the clips were cut from, and the transcript of the full recordingOutclip-map.md and clip-map.json
Around all of it
Working habits, and the skill that picks the next tool
- guardrails SkillThe safe-editing habits around the repo's rules.InNothingOutNothing; it changes how a session behaves
- video-tools-workflow SkillOne skill for the whole set: follows WORKFLOW.md and opens each tool's own SKILL.md at its step.InA recording, or a question about which tool does whatOutNothing of its own; it routes to the tool for each step
How they behave
Four rules the tools keep.
The plan is the deliverable
Every cutting and framing decision lives in a JSON file naming its source, its segments and its reasons. A render is disposable because its plan is not. A clip whose plan is lost is a clip nobody can explain.
Measured, not eyeballed
Geometry off pixels. Loudness off the file. Cut points off the waveform at 20ms, never off transcript timings, which are wrong by tens of milliseconds and will slice the middle of a word.
It proposes, you choose
Nothing here decides what is worth cutting. The tools narrow three hours to a page of candidates, and every one of them stays a candidate.
Nothing is published
These tools write files. No upload, no post, no push. What leaves an account is your decision, per action.
What it will not do
- Download from a platform. A recording arrives as a file from whoever owns it.
- Recognize faces. Nothing here knows who is in a frame. You look before anything goes out.
- Promise clean captions. A machine writes them, and names are the part it gets wrong most.
- Measure what performs. Clip length and framing follow editing habits with stated reasons. Nobody has measured what works best.