Writing ·
Building the 2026 reel
A reel for a Director-level reader: open on the strongest finished frame inside five seconds, then one measured system per tile with the number burned in. The rules, the inventory, and the site that carries it.
Most VFX reviewers watch muted. Recruiters say so in print, supervisors say so in interviews, and it changes what a reel for a studio innovation group has to do: prove, in silence, that the person builds systems that make many shots, not that they can make one pretty shot. So the 2026 reel has rules before it has a cut.
Open on the strongest finished client frame inside five seconds. That frame is the holographic play from Powerade x LSU, already public, so it carries no new clearance risk. It is also the poster on the home page of this site, chosen from a contact sheet of the master rather than from memory.
One idea per tile, with the number burned in. The AI matte bench at IoU 0.995 against a roto vendor, with the gate’s verdict on screen. The colour round trip with scopes and the drift figure. Slate nodes acting on real plates. The QC-reel builder producing a reel. Every tool tile is a real-time capture with a frame counter and a timestamp, no speed ramps, because a sped-up screen recording is the first thing a technical reviewer distrusts.
Provenance on every AI tile. Model, licence class, where the data stayed, whether the output reached final pixels, and the measured delta against the conventional path. The vocabulary is artist-directed and AI-assisted, never AI-generated, and nothing that resembles performance replacement belongs in it.
The inventory that came out of this was sobering in a useful way. The three “clearance-free” systems tiles are not clearance-free as the data sits: the bench ran on a client plate with a real coach in it, the colour pairs came from four client jobs, the depth runs from two more. So each gets re-run on my own footage with the same tools and the same numbers in the card before it appears on screen. The Systems pages carry the case studies now; the captures land with the re-runs.
The site is the companion. It was rebuilt this week on Astro so the case studies are content, not markup: every system entry has to carry at least one measured result and a provenance block before it builds. No iframe loads before a click, the fonts are self-hosted, and every route scores 99 or 100 on Lighthouse. The December cut replaces one id and one poster.
Update, 15 October: the capture plan. The seven case studies are approved and the media milestone has its list: twelve screen captures (two of them need no footage and were recorded first: the signed queue rejecting an unsigned intent, and the matte gate denying six injected defects) and a shoot of five short clips of my own, a talent walk, a clean plate, a colour chart with a log still, a handheld clip with depth, and a screen plate, so the bench, the colour round trip and the depth node run on footage I own before they appear on screen. The two recorded captures are live on their case-study pages; the rest land with the December reel, each one appearing here as its recording clears. The project pages get a before/after wipe as plate and comp frames clear: the first pairs come from a public breakdown reel, the rest wait on permissions.