Cutwise

An editor for people who hate editing

Cutwise started with a fairly simple question: why should cleaning up a screen recording take longer than recording it?

A typical software demo contains false starts, long pauses, repeated sentences, wrong turns through menus and a few seconds spent staring at a loading screen. None of that is difficult to fix in a traditional editor, but finding and cutting it all is tedious.

Cutwise is my attempt to remove most of that work. You give it the rough recording; it transcribes the speech, analyses what was said and what happened on screen, then proposes the parts that could be removed. You review those suggestions before it renders the finished video.

The aim is not to replace Premiere Pro, Final Cut or DaVinci Resolve. It is to make it possible to record a tutorial or product demo and get something publishable without needing them.

The edit starts with the recording

Cutwise never changes the original video. Every suggestion is stored as a decision against the source: remove this false start, shorten this pause, keep the second take or reject this edit.

The current pipeline can find long pauses, filler sounds, corrections, repeated takes, redundant speech and some visual mistakes. Every suggestion can be previewed before it is accepted, so the edit remains reversible and the AI never silently decides what survives.

Finding what transcription misses

One surprisingly awkward problem was filler words. Speech-to-text models often clean speech up as they transcribe it, which means the transcript can omit the exact "um" or "er" Cutwise is trying to remove.

Cutwise now runs a separate local acoustic analysis for those sounds. It combines overlapping transcription passes with audio boundaries to find likely fillers without changing the main transcript.

Visual analysis is optional. Cutwise samples a limited number of screenshots and uses them alongside the transcript to spot loading screens, wrong windows, repeated demonstrations and changes of section. Those screenshots are treated as evidence for a suggestion rather than permission to make an automatic cut.

More than jump cuts

Cutwise can add captions, section titles, text callouts, keyboard shortcut labels and subtle transitions. It also corrects loudness and applies gentle noise reduction before export.

The emphasis is deliberately restrained. Ordinary speech edits remain ordinary cuts, and extra polish is added only where it helps. Cutwise is supposed to make a tutorial cleaner, not cover it in animated graphics.

Local first, bring your own AI

Cutwise is a local desktop application for macOS and Windows. The video stays on the computer, Whisper handles transcription locally and FFmpeg processes and renders the finished file.

Features that need a larger model use a bring-your-own-key approach, with support for OpenAI, Anthropic and Gemini. When cloud analysis is enabled, Cutwise sends only the transcript or bounded set of sampled screenshots needed for that job, never the complete source video.

The interface is built in Flutter, with a local Python worker handling transcription and acoustic analysis. Suggestions become deterministic edit decisions that FFmpeg can render and the user can inspect. The important rule is simple: AI can suggest; Cutwise must make it easy for the human to disagree.

Current state

There is a working macOS build that can take a recording through the complete process:

Import → Transcribe → Clean up → Analyse → Review → Add polish → Render

It supports local transcription, filler detection, silence cleanup, semantic and visual analysis, transitions, callouts, captions, audio enhancement and MP4 export.

The macOS application is signed, notarised and working without development tools in its runtime path. I am now validating the Windows build and preparing Cutwise for a small private beta.