timbercut
Request a pilot

Revision Agent

The revision agent that shows its work.

Timbercut’s first product takes a revision round from start to finish. It reads every note against the project, writes an explicit edit plan, applies the changes it’s allowed to make, checks each one, and tells you exactly what still needs a person.

Run the sample roundSee the workflow
Spec · Revision AgentIn development
Takes in
A revision round: notes from every channel, plus the brief, rules, transcript, frames, assets and history.
Gives back
A proposed next version, its edit plan, a verification report and a revision audit.
People
Approve subjective, costly and rule-breaking changes, and sign off every round.
Today
The interactive demo runs a scripted simulation of one round on sample data.

01Context

It reads the project, not just the notes.

“Use the latest brand assets” only makes sense next to the brand kit, the asset bin and what happened last round. Timbercut builds that picture before it touches the timeline.

  1. Feedback

    Timecoded review comments, chat threads, email, notes docs and call notes.

    Sample round48 comments from 5 sources and 7 people, all on V7.

  2. Brief

    What was agreed, clause by clause, and who owns each clause.

    Sample roundApproved brief v2 §2.1: “Show the fold mechanism in one continuous shot of at least 4 seconds.”

  3. Brand, legal and delivery rules

    Guidelines, approved claims, legal notes and delivery specs.

    Sample roundBrand guidelines 2026.2 §4.2: “End-card logo lockup: between 12% and 18% of frame width.”

  4. Transcript and captions

    What’s said, what’s on screen, and which reads were approved.

    Sample roundVO end line: take 4, approved by the client in round two.

  5. Representative frames

    Key frames from each shot, so a note can be checked against the picture.

    Sample round00:00:10:20: the rider crosses on a red pedestrian signal.

  6. Asset metadata

    Bins, takes, tags, file specs and what’s actually present.

    Sample roundSH-05 take 4 is tagged “green signal” and runs the same length as take 2.

  7. Revision history

    Every earlier decision, kept in the Creative Decision Graph.

    Sample roundV5 enlarged the logo to 22%. V6 brought it back to 16% under the brand guidelines.

02Capabilities

Understand. Execute. Verify.

Twelve things the revision agent is designed to do, each shown with an example from the sample round you can run in the demo.

Understand01

Build the project picture and work out what each note means.

  • Reads notes in context

    Each note is linked to a timecode, shot, asset, brief clause or rule before anything changes.

    “the first few seconds still feel slow” → opening, SH-03 to SH-02

  • Consolidates repeats

    Notes asking for the same change become one request. Every source stays attached.

    C-01 + C-09 + C-19 → R-01, remove the price card

  • Resolves references

    Phrases like “Sarah’s comment from yesterday” are traced to the exact message.

    “Sarah’s comment from yesterday” → C-45, email, Tue 16:40

  • Finds trouble early

    Conflicts with the brief, dependencies between requests and missing files are flagged before editing starts.

    R-28 conflicts with brief §2.1 · R-43 waits on R-37

Execute02

Turn approved intent into production actions.

  • Writes explicit operations

    Every request becomes an operation you can read: what changes, where, by how much and why.

    remove_clip GFX_Price_1899 [00:00:21:00 → 00:00:23:12]

  • Applies what’s supported

    Routine, reversible changes run through connected production tools as a new version. The previous cut stays as it was.

    34 operations applied to a simulated V8

  • Asks before judgment calls

    Subjective, costly, legal or previously approved changes are proposed rather than applied, and routed to a named person.

    R-29 logo 16% → 18% · waits for Sarah Okafor

  • Supersedes without erasing

    When a later instruction replaces an earlier one, both stay on record with the reason.

    R-08 “raise the music” → superseded by R-33 “lift the VO”

Verify03

Check the result and account for every request.

  • Checks every change

    Each change is checked against the request behind it and the rules that govern it, with evidence attached.

    Lockup 16.0% within 12–18% (BR-4.2) · pass

  • Keeps blocked checks blocked

    If something can’t be verified, the report says so. A compromise is never recorded as compliance.

    9:16 end card present · blocked by R-37

  • Accounts for every comment

    Each comment ends the round with an outcome and a reason. That isn’t the same as every comment being done.

    48 of 48 accounted for · 34 completed

  • Writes the revision audit

    A record of the round you can keep or send: comments, requests, operations, checks and decisions.

    revision-audit.json · revision-audit.csv

03Approvals

Who decides what.

Timbercut goes ahead on routine changes and stops at the ones that belong to people. In the sample round, that’s 34 changes applied, 6 waiting for a named approver and 2 blocked.

When a change is… Example from the sample round What Timbercut does Who decides
Explicit, reversible, and inside the brief and rules “Typo in the super: ‘Foldes’ should be ‘Folds’.”Theo Brandt · R-05 Goes aheadApplies the change, then checks it against the request and the claims sheet. No one approves it first. Your team reviews the round before it goes anywhere.
Subjective, or the reading is inferred “The first few seconds still feel slow to me.”Sarah Okafor · R-02 AsksProposes one specific edit and labels the interpretation as inferred. Mina Cho, Creative DirectorCreative call
Changes something the client already approved “Use take 6 for the end line. Warmer read.”Mina Cho · R-35 AsksPrepares the change and holds it. Sarah Okafor, Brand LeadClient-approved element
Carries a cost or a license “Could we try the alt track, ‘Night Transit’? Feels more city.”Ava Lindqvist · R-25 AsksPrepares the swap and the retimed cuts, and holds them. Theo Brandt, ProducerCost and licensing
Changes a product claim “Change ‘Folds in 10 seconds’ to ‘Folds in seconds’. It’s a safer claim.”Daniel Reyes · R-30 AsksPrepares the new wording and holds it for legal review. Priya Nair, Legal CounselLegal review
Repeats a decision that was reverted “Make the logo bigger on the end card.”Daniel Reyes · R-29 AsksShows what happened last time and proposes the largest size the rule allows instead. Sarah Okafor, Brand LeadBrand decision
Conflicts with the approved brief “Cut the fold demo. It slows the whole spot down.”Daniel Reyes · R-28 BlocksStops, explains the conflict and sets out options. A shorter compromise still needs an exception. Sarah Okafor, Brand LeadBrief owner
Needs a file nobody can find “The new vertical end card from the brand team should go on the 9:16s.”Sarah Okafor · R-37 BlocksLists every place it looked and drafts a request for the file. Anything that depends on it stays blocked. Sarah Okafor, Brand LeadAsked for the asset

Whatever the policy, a person signs off the round before anything goes to the client.

04Outputs

What comes back at the end of a round.

Not just a new cut. Everything you need to trust it, question it and send it.

  1. 01

    A proposed next version

    V8 is built alongside V7. Nothing is overwritten, and every change can be traced back.

  2. 02

    The edit plan

    Every request as a readable operation, with its sources, its reasoning and the rules it touches.

  3. 03

    A verification report

    Checks and evidence for each change, plus round-level checks for the brief, brand and legal rules.

  4. 04

    A short decision queue

    Each approval names the person who owns it and the reason it needs them.

  5. 05

    Drafted replies

    Requests for missing files and explanations for declined notes, ready for a person to edit and send.

  6. 06

    The revision audit

    The full record of the round as JSON or CSV, for your files or for the client.

05Questions

The questions producers ask first.

Straight answers, including what isn’t built yet.

Ask us something else

What happens when comments conflict?

Timbercut doesn’t pick a winner by default. When two notes disagree, it looks at who said what, when, and with what authority. A later instruction from the brand lead in a review call, for example, supersedes an earlier request to raise the music, and both stay on record.

When a note conflicts with something agreed, such as the approved brief, the request is blocked and the options are set out for the person who owns that agreement. A compromise, like shortening a required shot below its minimum, is still treated as an exception that needs approval.

What needs approval?

Anything subjective, anything where Timbercut had to infer the meaning, and anything with cost, legal or contractual weight: changes to elements the client already approved, new product claims, music licensing, and changes that break the brief or brand rules. Each one is routed to a named person with the reason it needs them.

Your team decides where that line sits during a pilot. Nothing goes to a client automatically.

What happens when an asset is missing?

The request is blocked rather than guessed at. Timbercut lists where it looked, drafts a message asking for the file, and keeps anything that depends on it blocked as well. In the sample round, the vertical versions can’t pass verification until the new 9:16 end card exists.

How are decisions tracked?

In the Creative Decision Graph: who asked for a change, where they said it, which version it touched, which rule applied and what happened next. Later decisions supersede earlier ones without erasing them, and inferred patterns are labeled so they’re never mistaken for rules.

What can I see working today, and what’s planned?

Today, this site runs an interactive simulation of one revision round on fictional sample data. It runs entirely in your browser. It doesn’t edit video, call an AI model or connect to any tool.

The revision agent itself is in development, and connections to review, chat, storage and editing tools are planned. The integrations page lists each one and its status.

Does it replace editors?

No. It takes on the clerical part of a round: collecting notes, working out what they mean, making routine changes and checking them. Editors and creative directors keep the craft and the judgment, and they review everything before it leaves the building.

Which AI does it use?

Timbercut is designed to use Claude, Anthropic’s AI model, as its reasoning and orchestration layer: reading the project context, interpreting notes, planning operations and deciding which tools to call. The architecture is described on the integrations page.

Nothing on this website calls Claude or any other model. The demo is scripted. Timbercut isn’t affiliated with or endorsed by Anthropic.

What happens to our footage and documents?

Pilots are scoped one project at a time. Where material is stored, who can access it and how long it’s kept are agreed with you in writing before anything starts. We don’t claim any security certification.

Watch it work a round of notes.

48 comments, 5 sources, one conflict with the brief and a missing file. The demo takes about a minute.

Run the sample round Request a pilot