Noirpost — overnight post for technical streams

The problem, in one second of footage

Auto-croppers throw away the thing your clip is about

Clipping tools are built for podcasts. They find a face and crop to it. Point that at a stream where the content is a terminal, an IDE, a firmware dump or a dashboard and you get a clip of nothing. Below is the same second of the same stream, cut two ways.

A 9:16 crop of a 16:9 stream frame: terminal text runs off both edges and the speaker is outside the frame. — Cropped to 9:16 the usual way

Every line of output cut in half at both margins. The speaker is gone. No hook, no captions, nothing telling a scrolling viewer what they are looking at.

The same moment as a finished vertical Short: hook line at the top, the speaker framed on his own face, the terminal readable below, karaoke captions and a branded bar. — The same second, ours

Hook line up top. The speaker framed on his own detected face. The terminal kept whole and readable underneath. Word-by-word captions, sound design, your bar and handle.

What actually happens to your stream

Eleven stages. Machines do the measuring, people do the judging.

Layout, crops, geometry and timing are deterministic — the same stream cuts the same way every time. A language model is only allowed editorial calls, and every one of its answers is clamped by code before it reaches a render. Then a person watches the result.

01Pull the streamA public URL is all we need. No channel access, no credentials.Code

02Transcribe onceFive and a half hours, word-timestamped, in about a hundred seconds. Nothing is transcribed twice.Code

03Find the momentsReveals, strong opinions, things that broke, things that worked. Shout-outs and dead air thrown out.Model

04Cut the candidate list downRoughly thirty-five become twenty-five. Hooks get rewritten where they read generic.Human

05Locate the words in file timeKeyframe seeking lies by up to ten seconds, so the moment is found by matching speech, never by the clock.Code

06Decide the editUp to four ranges per clip. The first delivers the hook, the last lands on the strongest line.Model

07Snap every cut to real speechSentence ends first, then pauses, then clauses. Ranges that overlap are merged; dead air is dropped.Code

08Read the frameScreens and cameras separated by saturation, faces found per panel, terminals checked for readable text by OCR.Code

09Choose the layoutScreen-led, two-up, reaction or solo — picked per clip. Empty screens are never shown.Code

10Captions and sound designTechnical jargon repaired, then two to five effects placed so their impact lands on the beat.Model

11Watch every clip, then shipStart to finish, by a person. Weak clips are rebuilt or killed before you ever see them.Human

Code — deterministic, repeatable, measured Model — editorial judgement, clamped by code Human — taste, and the final no

Real output

The layout is chosen per clip, not stamped on every clip

Four untouched frames from delivered batches. Each got a different composition, because each moment had a different thing worth looking at.

Short: a full-width terminal above a strip showing the streamer's face.

Screen leads

Dense terminal output gets the panel. The face drops to a strip.

Short: a single close-up of the streamer, no screen content at all.

Nothing on screen

When the screen holds nothing, it isn't shown. One close-up instead.

Two people

Each face framed on its own face, not on a single tracked speaker.

Short: a narrow screen strip, a large facecam, and a chat message pulled into frame.

Chat in frame

When the moment is a chat message, the chat becomes part of the clip.

Elapsed time from the moment your link lands