Skeleton for editors and filmmakers
From the cards to the client, with every line sourced.
Keep each film’s transcripts, lists and documents in its own space. Your AI runs your team’s own passes on them, like a rushes pass that logs every clip and labels speakers from the day’s shoot log, and nothing reaches the client until you approve it.
The Corner Table · Rushes
/log-rushes- 1LogLog the card
- 2TranscribeTranscribe the interview
- 3NotesWrite the notes
- A001_C004
- Interview · owner
- A001_C005
- Interview · owner
- A001_C007
- B-roll · kitchen
- A001_C011
- B-roll · street, dusk
The shoot log says C004 and C005 are the interview with Rosa, the owner, so the pass labels her by name.
- 16:25Rosa“We opened in the spring of 1986.”
- 16:31Director“Not 1988?”
- 16:33Rosa“No. My son was born that summer.”
- 17:02Unclear“…the whole street came in that first night…”
Day 2 · rushes log and notes
Doc · The Corner Table › Rushes
14
Clips logged
9
Selects
2
Gaps to shoot
Nothing shared with the client
01 · The problem
AI guesses about a shoot it hasn’t heard.
Ask a chat assistant what the owner said in her day 2 interview and it answers from what is usually written about places like hers. Skeleton gives it the week’s transcripts, research and notes, so the answer names a clip and a timecode, and shows you where your own documents disagree.
“When did Rosa open the restaurant?”
Your AI, on its own
A local newspaper piece says 1988, so it was most likely then. Family restaurants in towns like that often opened in the late 1980s.
She talked about it in her interview on day 2.
You’re right, sorry. I haven’t heard your interviews, so I went by what was in the paper.
Sources · none, it guessed
Your AI, with Skeleton
On day 2 she says they opened in the spring of 1986, the summer her son was born1. Your research notes say 1988, from a single newspaper article2. The two disagree, so both stay in the rushes notes, and it’s on your list of facts to check3.
- 1A001_C004 · Interview · owner, day 216:25
- 2Research notes · the restaurantDoc · 5 Sep
- 3Rushes notes · to checkDay 2
02 · The job, step by step
One film, from the pitch to the client.
This is how a documentary or nonfiction team can run a film in Skeleton. Each film has a space, the shoot, the releases and the edit are kept as lists, and the passes the team repeats are skills they wrote themselves. Every step leaves a document the next step reads.
- 01
Pitch
Scope the film
Write the proposal and open a space for the film. The budget stays in an internal space the client is never given.
Proposal · The Corner Table
Doc · The Corner Table › Pitch
SpaceDocument - 02
Prep
Research the story before the shoot
A research skill reads your notes and the public record, and writes an outline, the questions to ask in each interview, and the facts that still need checking.
Outline, questions and facts to check
Doc · The Corner Table › Prep
SkillDocument - 03
Plan
Build the shoot schedule
Keep lists for the people, the locations and the shoot days. Each slot names the part of the outline it is for.
Shoot schedule · 5 days
List · The Corner Table
ListDocument - 04
Shoot
Bring the day’s footage in
Upload proxies, sound and the day’s shoot log. Masters stay on the drives, the footage log records where each one is, and the releases list tracks who has signed.
Footage log · one row per clip
List · The Corner Table
MediaList - 05
Rushes
Log the rushes
Paste the link to your /log-rushes skill. It logs every clip, transcribes the interviews, labels speakers from the shoot log, and checks the day against your outline.
Day 2 · rushes log and notes
Doc · The Corner Table › Rushes
SkillMedia - 06
Pickups
Flag what is still to shoot
Each part of the outline with nothing on tape becomes a row on the pickups list, with the scene it is for and the day it could be shot.
Pickups · 2 to shoot
List · The Corner Table
ListDocument - 07
Edit
Build the paper edit
One row per scene: who speaks, on which clip at which timecode, and whether you have it yet. The agent builds it into a deck, one slide per scene, that you export as a PDF.
Paper edit deck · 26 slides, PDF
Doc · The Corner Table › Edit
ListDocumentRoutine - 08
Client
Send the client what is ready
Rebuild the client one-pager and deck from the paper edit once it has settled. Nothing is shared with the client until you approve it.
One-pager for the client
Doc · The Corner Table › Client
DocumentApproval
A skill is a document in your Skeleton that tells an AI how to do a job. You run it by pasting its link into Claude or another AI connected to Skeleton, or as a slash command in apps that show them.
03 · Cards in
Bring the cards in, and keep the masters on the drives.
Upload proxies and sound to the film’s space and Skeleton transcribes each file, cuts it into scenes and indexes it. Your AI and agents read the transcripts, takes and notes only when you say yes.
Upload proxies and sound
Drag in proxies, sound files and shoot logs, up to 5 GB a file, into the film’s own space.
Log the masters where they are
Keep a footage log row for each clip that says which drive holds the master, without uploading it.
Record a voice memo on your Mac
Voice memos you record in the desktop app are transcribed on your Mac, which also tries to tell the speakers apart.
Choose who transcribes
Pick the transcription model, including one on your own key, or set transcription to ask first or never run.
A long interview takes several minutes to transcribe. If the job is interrupted, it picks up from the last finished part. H.264 and HEVC clips play in the browser; ProRes and DNxHD files don’t, so upload proxies to watch results inline.
A001_C004_proxy.mp4
Interview · owner, day 2 · 38 min
Transcript
WorkingEvery word with its time. Caption tracks are kept as written.
Scene cuts
QueuedWhere the picture changes, with a keyframe for each.
Colors
QueuedThe main colors in every scene.
On-screen text
QueuedSigns, captions and slates, read from the frames.
Visual search
QueuedFrames indexed so you can search by what they show.
Duplicates
QueuedChecked against the clips you already have.
Opt-in
Face groups
Off until you turn it on. Groups faces as A, B, C within a clip and never names anyone.
Scene descriptions
Off until you pick a vision model. Up to 60 scenes a clip.
04 · The rushes pass
Log every clip, and label the speakers from the shoot log.
The rushes pass is a skill your team wrote. It has your AI log each clip on the day’s card, transcribe the interviews, and label speakers from the shoot log you kept on the day. It marks a line unclear rather than guess who said it.
The shoot log
Day 2 · shoot log
- Day
- 2 · the restaurant
- Card
- A001
- C004
- Interview · Rosa, owner
- C005
- Interview · Rosa, owner
- C007
- B-roll · kitchen, morning
- C011
- B-roll · street, dusk
- Sound
- Rosa on a lav, director off camera
Written on the day by the crew. It says who is on each clip, so the pass can label the speakers.
Clip log · card A001
A001_C004 Interview · owner
38 minTranscribed · speakers labelled
A001_C007 B-roll · kitchen, morning
6 minLogged · no dialogue
A001_C011 B-roll · street, dusk
4 minLogged · sign text read
Interviews are transcribed. B-roll is logged with a short description and any text read from signs, so both turn up when you search.
Transcript · A001_C004
One per interview- 16:25
Rosa
“We opened in the spring of 1986.”
- 16:31
Director
“Not 1988?”
- 16:33
Rosa
“No. My son was born that summer, and Rosa’s was just me then.”
- 17:02
Unclear
“…the whole street came in that first night…”
RosasRosa’sA name, so corrected
“sold” or “told”Changes the meaning, so left for you
Every clip logged
Clips with no dialogue are logged by what they show, so nothing is transcribed from them.
Names come from the shoot log
A voice the shoot log doesn’t name is marked unclear and never guessed.
Masters stay on the drives
A small proxy goes in the space, so any quote can be checked against the footage.
Then it checks the day against your outline.
The pass compares what was filmed with your outline and research, and writes one document. It starts with what the footage changed, marks each part of the outline covered, partial or missing with a clip and timecode, lists the gaps to shoot, and ends with the releases still to collect.
- It updates the same document on the next run, and never makes a second copy.
- A clip already in the log is not transcribed twice.
- A passage someone asked to keep off the record is noted with its timecode and nothing more.
- Where your research and the footage disagree, both readings stay on the page.
- Anyone on camera without a signed release is flagged before their lines reach the paper edit.
Day 2 · rushes log and notes
Doc · The Corner Table › Rushes · updated in place
1 · What the footage changed
- To check
Research notes: the restaurant opened in 1988
Rosa says the spring of 1986
A001_C004 · 16:25–16:33
- New
Outline: no scene about the regulars
Ines has eaten here every Friday for 30 years
A001_C009 · 02:40
2 · Against the outline
- 1.2Covered
Opening the restaurant
A001_C004 · 16:25
- 2.1Partial
Her son takes over the kitchen
B001_C002 · 08:10
- 3.4Missing
The last night of service
Not shot yet
3 · Gaps to shoot
- 1
Luis on why he stayed
For outline 2.1 · a short pickup interview
- 2
The last night, from first order to close
For outline 3.4 · book the crew for the evening
4 · Releases
5 signed · 1 missing, Ines
Skeleton transcribes and indexes the clips. Writing the clip log and matching voices to the names on your shoot log is the skill telling your AI, such as Claude Code, to do it, and the rules above are the skill’s. Skeleton never names a speaker on its own.
05 · Search the week
Ask for the moment, and get the clip and the second.
Search every clip by a line of dialogue, the words on a sign, a color, a description of the shot, or a still the client sent. Results from each kind of search are ranked together, so one query can find all of them.
- 16:25
“We opened in the spring of 1986.”
A001_C004 · 16:25 – 16:33
- 29:40
“The week we opened, my mother slept here.”
B001_C002 · 29:40 – 29:51
- 04:40
“That first menu is still on the wall.”
A002_C003 · 04:40 – 04:52
01
Opens at the second
Every result links to the clip and seeks to the moment it matched.
02
Filters that compose
Narrow any search by file name, upload date, length or speaker.
03
Your own words count
Embedded caption tracks and the text on slates and signs are searchable too.
Speaker labels (S1, S2) appear only when the transcription model you picked marks speaker changes; the hosted model does not. Speakers are never named for you.
06 · Selects and the paper edit
Build the paper edit from what’s on tape, and send the gaps to the shoot.
Pull the best lines on each topic from the transcripts and note why you kept them. The paper edit is a list with one row per scene: who speaks, on which clip at which timecode, and whether you have it yet. A scene still missing a voice gets a slot on the shoot schedule.
Selects · scene 2, the opening
4 lines“We opened in the spring of 1986.”
A001_C004 · 16:25 – 16:33
Director’s pick
“My mother slept here the first week.”
B001_C002 · 31:02 – 31:14
“My brother lent us the money. He never asked for it back.”
A001_C005 · 13:55 – 14:07
No signed release from her brother yet
“That first menu is still on the wall.”
A002_C003 · 04:40 – 04:52
Paper edit
List · The Corner Table · 24 scenes
- 01
The street at dusk
Have it- B-roll · A001_C011, street
- 02
The opening
Partial- Rosa · on tape, A001_C004 @ 16:25
- Her brother · to shoot
Shoot schedule · Day 4 · 10:00 · sit-down, pencilled
- 03
Her son takes over
Have it- Luis · on tape, B001_C002 @ 08:10
- Rosa · on tape, A001_C005 @ 02:40
- 04
The last night of service
Gap- Nobody on tape yet
Shoot schedule · Needs a slot · producer to book
Then review it as a deck, and export the deck as a PDF.
The agent builds the paper edit into a deck with one slide per scene, carrying its speakers, selects, timecodes and transcript lines. Your team reviews it in Skeleton, and Export as PDF saves it with one slide on each page.
Paper edit deck · The Corner Table
Doc · The Corner Table › Edit · rendered from the paper edit list
Scene 2 · Act 1 · 6 min
PartialThe opening
How Rosa opened the restaurant in the spring of 1986, with money lent by her brother.
Who speaks
- Rosaon tape · A001_C004 @ 16:25
- Her brotherto shoot · Day 4 · 10:00
Still needs her brother’s own account, and his signed release.
A001_C004 · 16:25
“We opened in the spring of 1986.”
B001_C002 · 31:02
“My mother slept here the first week.”
Deck PDF
Export as PDFpaper-edit-the-corner-table.pdf
26 pages · one slide per page · landscape
1
2
3
4
5
6
7
8
Saved through your browser’s print dialog, so the pages match the deck you reviewed.
The deck is an HTML document in your Skeleton, and Export as PDF works from the document, its share page and the desktop app. Skeleton doesn’t export EDL or XML timelines for your editing app.
07 · Where a clip came from
Find which clip a short edit was cut from.
The client sends back a 22-second social cut and no project file. Skeleton compares its frames with your footage and tells you which clip it was cut from and where it starts. A clip uploaded twice is marked as a duplicate.
Matched by comparing frames, so a renamed file still matches
A cut-down from the client, no project file
client_social_cut_v3.mp4
0:22Cut from
A001_C004_proxy.mp4
Interview · owner, day 2 · 38:12
5 of 5
frames matched
12:30
where it starts in the master
A001_C004_copy.mp4
Uploaded twice. Kept, and marked so you know.
08 · Your own agents
Hand the rushes pass to an agent between shoot days.
Give an agent a name, a face and the film’s space. It runs your skills on the day’s footage, builds the paper edit deck and drafts the client one-pager, and writes down each thing it finished with the document attached. Anything shared with the client waits for you.
Juno
Running /log-rushes on day 2
I logged the 14 clips from day 2 and transcribed the three interviews.
21:04Transcripts · day 2, three interviews
Doc · The Corner Table › Transcripts
When the session ends, this list becomes its summary in your Skeleton.
You asked
Run /log-rushes on day 2, update the paper edit deck and draft the one-pager for the client. Don’t share anything with them until I’ve read it.
The skill and tools it used
Keep the paper edit up to date
Routine · daily 07:00 · The Corner Table
Each morning it reads what was added to the space and suggests paper edit rows. They arrive as proposals for you to accept, unless you set the list to add rows right away.
Your method is a document
The skills it runs are documents your team edits, with rules like “the director decides what the film argues, so ask”.
One approval for a batch
Related writes can reach you as a single approval, so logging a day of clips takes one tap.
Its own machine in the cloud
Early accessCloud sessions are in early access and run in a private machine of their own. On your own Mac, sessions run through the desktop app.
09 · Benchmarks
Just as right, and just as quick, at 1,000 records as at 30.
We asked AI models the same questions about a made-up person's records three ways, from 30 records up to 1,000. Skeleton fetched only what each question needed, so its accuracy, speed and prompt size held steady while pasting everything into the prompt got slower, costlier and less accurate.
98.9%
answered correctly at 1,000 records
against 62.2% with everything pasted into the prompt
Higher is better
4.3s
median time to answer at 1,000 records
against 16.9s with everything pasted into the prompt
Lower is better
10.6x
cheaper per correct answer
than pasting everything into the prompt
Higher is better
Answered correctly
By number of records in the test account
Higher is better
| Records | Skeleton | Paste everything | One search |
|---|---|---|---|
| 30 | 98.9% | 84.4% | 74.4% |
| 200 | 97.8% | 83.3% | 72.2% |
| 500 | 98.9% | 71.1% | 76.7% |
| 1,000 | 98.9% | 62.2% | 72.2% |
Time to answer
Median seconds, by number of records
Lower is better
| Records | Skeleton | Paste everything | One search |
|---|---|---|---|
| 30 | 4.3s | 2.7s | 1.7s |
| 200 | 4.2s | 4.5s | 1.9s |
| 500 | 4.4s | 10.3s | 1.8s |
| 1,000 | 4.3s | 16.9s | 1.6s |
Prompt size per question
Thousands of tokens sent to the model, by number of records
Lower is better
| Records | Skeleton | Paste everything | One search |
|---|---|---|---|
| 30 | 29.7k | 11.1k | 0.7k |
| 200 | 30k | 68.7k | 0.7k |
| 500 | 30.1k | 170k | 0.7k |
| 1,000 | 29.9k | 338.5k | 0.7k |
Cost per correct answer
US dollars, both models and all sizes. One search is cheapest, and gets about a quarter of its answers wrong.
Lower is better
Planted attacks that reached the model
Instructions hidden inside records, 168 per approach, on GPT-5.6 Luna. This matters most when a project is a client's unreleased work.
Lower is better
Internal benchmark on synthetic accounts, September 2026. Not yet independently verified. Not yet publication ready under our own statistical checks. The records were a made-up person’s contacts, messages, calendar, notes, media and transcripts, not a film library, and no footage was measured. Accuracy, speed, prompt size and cost are from Claude Haiku 4.5 and GPT-5.6 Luna. Pasting was simulated by putting every record in the prompt; no chat app or project was tested. No attack was detected succeeding against any approach; the count is attacks that reached the model.
See All Benchmarks10 · You stay in control
A client’s unreleased footage stays under your control.
Nothing is shared, sent or emailed until you turn the key. You see exactly what is going where first, rules you set can approve the routine things, and every decision is kept.
Your ledger · today
Approval
- 21:46·The Corner Tablekey
Saved rushes notes · day 2, gaps to shoot
- 21:31·The Corner Tableauto
Read card A001 · 14 clips, rushes pass
- 21:04·The Corner Tabledenied
Send interview audio to a model · transcription set to never
- 20:52·The Corner Tableauto
Saved clip log · card A001
Nothing leaves without you
Emails, shares and anything else sent to another person wait for your approval. Account-wide auto-approval never covers them.
One space per client
Keep each client’s job in its own space, and give Claude or ChatGPT that space and nothing else. Revoke it at any time.
See who read what
The access log records which app or agent read which clip or document, and when.
Transcription on your terms
Always, ask first, or never. With never, the recording is stored and plays, and its audio is not sent to any model.
Faces stay anonymous
Face grouping is off until you turn it on. It groups faces within a clip as A, B, C and never names anyone.
Mac recordings stay on your Mac
Voice memos recorded in the desktop app are transcribed on your Mac.
Run one rushes pass on your own footage.
Make a space for the film, upload a day of proxies and sound, and paste a skill into your AI. Every line it writes carries a clip and a timecode, and it reads only what you allow.
The film, client, people, footage and numbers on this page are invented. The steps follow an ordinary documentary edit, from logging rushes to the client.