The boring half of editing

Nine tenths of editing is not creative work: moving pieces around, subtitles, levelling audio, and finding the moments worth keeping inside hours of recording. Everything here takes that part off a person and leaves the decisions.

Built out of my own years as an editor rather than out of an idea of what editing is like. That is also where the limit comes from, and it is said out loud here: an editor does not need a tool that decides for them — they need one that does the mechanics.

Zhnyva — a clip farm that runs itself

A clip farm on a server: it walks the clip-campaign marketplaces itself, picks the moments with a model, cuts them with subtitles and hands over a batch to publish, already marked as an ad.

Where it stands: Running on my own server since 27 Aug 2026, with a Telegram bot as its face. 332 checks and a lock built out of 52 deliberate breakages. The screenshot shows the counters row only: below it sit other companies brands, and that is not mine to show.

Python
yt-dlp
ffmpeg
Groq
+2

Full storyBehind a password

A catalogue of 237 Premiere plugins — other people's, not mine

A searchable reference to 237 Premiere plugins and presets — other people’s, not mine — with a column no store has: will this still run in the 2026 version.

Where it stands: Assembled on 13 Aug 2026 and checked. Useful reference work; the plugins themselves belong to their authors.

HTML
search and filters

Full story

Premiere Panel — auto-editing panel for Adobe Premiere Pro

An extension inside Premiere Pro that takes over the mechanical half of editing: finding the moments, planning the camera cuts, building the subtitles, levelling the sound.

Where it stands: Built (parts 1-7). On a real 17.5-minute 4K interview the engine produced a 15-step edit plan and marked 11 steps as ones it can perform itself; the plan has never been applied inside a live Premiere session.

Python (stdlib only)
UXP panel
Ollama (local)
faster-whisper

Full story

Content Factory — autonomous short-video pipeline

A short-video pipeline that walks the whole way from a trend to a publication on its own — except for one place where a human says yes.

Where it stands: Ran daily from 2 Aug 2026; deliberately paused on 10 Aug 2026.

Python
n8n
local image and voice models
Telegram

Full story

Local video generation on my own GPU

Video generated on my own graphics card instead of a paid cloud service — an experiment costs a second, not an invoice.

Where it stands: Working; used as the render stage of the Content Factory.

Python
RTX 5060
local diffusion models

Full story

Prisma — translate anything on your screen

A screen translator: select text in any window and the translation appears beside it. Plus text from images, PDFs and video, voice input, and translating a whole window at once.

Where it stands: Working, 121 automated checks.

Python
OCR
local and free cloud models

Full story

Ear — meeting transcription that does not quietly lie

Meeting transcription into text and notes — built because of one finding: a normal recogniser that gets the language wrong quietly translates instead of transcribing.

Where it stands: Working. The recordings are other people talking.

Python
faster-whisper
local models

Full storyNo screenshots: they would show other people's data

Almost all of it is computed on my own machine: transcription, choosing the moments, generation — with local models. That is not ideology but arithmetic: video work happens in hundreds of attempts, and paying per attempt turns the search itself into a budget line.