Distill, Then Critique

A video on lighting runs about 1 hour and carries roughly 9,000 spoken words. Strip the intro, the 3 sponsor reads, the recap and the sign-off and what is left is around 35 minutes of instruction that actually concerns the craft. Paste the whole 1 hour in and the machine reads all of it, because it does not drift at minute 52 and it does not stop taking notes when the diagram on the whiteboard gets messy. The mistake is treating the transcript as the deliverable. A transcript is ore. It is heavy with the right material and lighter on signal than anyone expects on the first read, and the fix is not a better video. The fix is a refinery step between the video and the work. 2 transcripts went in. 1 covered lighting and 1 covered camera loops. What came back was not 2 summaries. It was a working skill set, and the output quality moved on the very next job. Nothing about the camera changed and nothing about the crew changed. What changed is that the knowledge stopped being something 1 person had watched and became something 2 people could execute. A transcript is not the deliverable. A transcript is ore.

The method is 4 steps and every 1 of them is mechanical. 1 is to find a video that covers the craft in detail, made by somebody doing the work rather than talking about it. 2 is to copy the transcript, which the platform will hand over as full plain text in 1 click. 3 is to paste it in and ask for the whole thing to be pulled together, which is the instruction that produces an actionable skill set instead of a summary. 4 is to read what comes back and use it on the next job. The gap between a summary and a skill set is the gap between knowing that light has 3 properties and knowing where to place 3 lamps for a far-side key at 45 degrees. A summary tells you what was said. A skill set tells you what to do at 7am when the lamp kit is 1 head short and the sun is already 20 degrees off the window. 4 steps, 1 sitting, and the material comes out as procedure. 4 steps, 1 sitting, and the material comes back as procedure.

Fed properly, the distill arrives structured. A numbered set of rules, an ordered sequence of setup steps, a short list of failure modes, and the 1 or 2 numbers that make the whole thing repeatable. The lighting distill carried the 45 degree far-side key, the 2 to 1 ratio between the hot side of the face and the shadow side, and the 3200 to 5600 Kelvin pair that decides whether the frame reads warm or clinical. The camera-loop distill carried the loop lengths, the 180 degree rule, and the 3 planes that have to sit inside 1 frame for the shot to read as depth rather than flatness. That is the shape of a usable distill. Rules that can be counted, steps that can be followed in sequence, and numbers that can be checked on set. Anything that comes back as prose about how important it is to understand light has failed the step, however well it is written. The test is 1 question. Can somebody open this on set at 6am with 20 minutes to rig and follow it. If yes, the distill is done. If no, the paste was too broad and the ask was too vague.

Here is the step almost everyone skips. Take the distill, hand it to a second model, and ask 1 question: is there a better way to get this from the model? The second pass is a critique, not a rewrite. Where the first pass turns a transcript into rules, the second pass looks at the shape of the ask itself. Are the 4 steps in the right order. Is 1 instruction quietly doing 2 jobs. Can a rule be cut to 1 line without losing the decision inside it. Is a number being asserted where a judgement belongs. That framing is why the question is worth asking out loud. Most people ask a model to do more. The useful question asks the model to look at how it is being asked, which is 1 level up from the content and 1 level closer to where the waste lives. The second pass is short. It reads 1 document and returns a tightened version of the same method. What it cannot do is decide for you whether the rule is right, because the rule still has to survive contact with a set. The second pass is a critique, not a rewrite.

A transcript is not the deliverable. A transcript is ore.

Compare 2 ways to find the same defect. On set, a mistake in the lighting setup costs a shoot day. 8 hours of crew time, 1 location held for a further 24 hours, and a re-rig that burns the 2 hours of usable daylight the scene needed. Add travel, and 1 wrong line in a setup sheet can cost 3 days and a full camera package. On text, the same defect costs 1 paste and 40 seconds of reading. The critique model never leaves the desk, never holds a permit, and never bills a day rate, so the correction is effectively free at the moment it is made. That asymmetry is the whole argument. A defect caught at the text stage costs minutes and a defect caught at the render stage costs days, so every cheap pass should happen before the first expensive 1. 1 render per critique and no render before the critique is a rule that pays for itself on the very first job it saves. A mistake on text costs 40 seconds. The same mistake on set costs a shoot day.

4 checks before a distill goes into the workflow, and none of them is a glance. Read it back and confirm the rules are countable, because a rule that cannot be counted cannot be checked on set. Then confirm every step is followable in sequence, because 1 step that assumes knowledge the crew does not have will stall the rig. Then confirm the numbers are the numbers, since a transcribed figure like 3200 Kelvin becomes wrong the moment it is rounded in a summary. Then run the whole thing past the second model once more with the same 1 question, because a tightened method can be tightened again. Here is the arithmetic behind all of it. Built companies across 12 countries. Restructured a EUR 75 million industrial group. Deployed 210 energy systems across Africa and Asia. Took a company through a Frankfurt IPO. 16 years of field work from rural Nigeria to Fukushima. Entrade built 200 wood-gasification machines across the UK and Europe. Signed a EUR 20 million personal guarantee on a company that folded on Christmas Day 2008, and lost savings in the 2013 Cyprus bail-in. 393 pages in paperback and Kindle, published 8 April 2026. None of it arrived as 1 insight. All of it arrived as procedure, tightened by a second pair of eyes before it was expensive to be wrong. The probability that 1 video watched end to end produces a working skill is close to 0. The probability that the same video, distilled and then critiqued, produces 1 is close to 1. Most people price the first at 100 and the second at 0, which is why they finish the video and still rig the light the old way. Hand it to a second model and ask 1 question: is there a better way to get this from the model?