The Install Proves The Machine Before The Film

The folder arrives complete and it still does not run. Every copied script carries the absolute path of the machine it was authored on, so 4 scripts point at a directory that does not exist on the new computer. Step 0 is 1 one-time edit per script: repoint the path variables at the top. The stills generator carries 1, the video generator carries 1, the assemble script carries 2, the grade script carries 1. That is the whole fix, and it is the step that costs an afternoon when it is skipped, because the failure surfaces 20 minutes later as a missing directory rather than as a path error. A future tidy-up would centralise those into 1 config file, and the scripts are proven as they stand, so the edits stay minimal. We built 200 wood-gasification machines across the UK and Europe on the same discipline. The machine that arrives with someone else's settings is not installed. It is unpacked.

The prerequisite list is 3 items: Python 3.9 or newer with the imaging library, ffmpeg for trim, concat, mux and audio mix, and the generation CLI. All 3 are standard on most development machines, and none of them should be assumed present, because 1 missing binary fails at the last step of a 12 step pipeline rather than the first. So each one is verified by 1 line that returns a version: the ffmpeg version line, the library version print, the CLI version. 3 commands, 10 seconds. A prerequisite is a claim until something prints a number for it. I signed a EUR 20 million personal guarantee against a plan that assumed every future moment would arrive on schedule. The company folded on Christmas Day 2008. What I kept from it is the version check.

3 keys: 1 voice key that is required, 1 key covering stills and video, and 1 optional fallback for stills. All 3 get exported into the shell profile and pasted into nothing. A key typed into a chat window is a key in a transcript, and a key in a transcript is a key that will be rotated. The template file is copied to a local secrets file and filled in there. Then the CLI is authenticated once, and never again on that machine. The same 3 export lines sit in the shell config, which means the pipeline starts from a clean terminal with 0 setup per run. The entire secret surface of a film pipeline fits in 3 exported variables. 12 departments and 1 operating standard was the version of this at a EUR 75 million industrial group. The smaller the surface, the fewer places a machine can be wrong.

Step 3 is the part that matters, and it is 3 artifacts rather than a film: 1 voice line written to an audio file, 1 still, 1 clip. Each one is asserted by a printed byte count, and each one is opened before anything downstream is touched. The voice test prints the file name with the byte count attached. The still is 1 single 720x1280 jpg on disk. The clip is 1 single 768x1364 file. Only after 3 files exist and 3 byte counts look sane does a full film get attempted. The reason is arithmetic. A 12 step pipeline has 12 places to fail, and 3 smoke artifacts separate the generation stack from the assembly stack before either one is asked to do 40 shots. 16 years of field work, from rural Nigeria to post-conflict Serbia to Kazan to Fukushima, taught the same ordering. Prove the instrument first. Then measure, then build, then ship.

The machine that arrives with someone else's paths is not installed. It is unpacked.

Production is a sequence, not a menu. 1 script holds 1 voice line per shot, and the scripts run in order: voice, stills, video, assemble. Writing the script first is what makes the order mechanical, because every downstream step reads the shot count instead of guessing at it. The assemble script is idempotent, so a failure in the middle is a re-run rather than a rebuild, and the frame count handles its own off-by-one without a manual fix. There is a library of these lines in the prompts folder for the days when the structure is right and the words are not. 210 energy systems across Africa and Asia were deployed with the same shape: 1 written order, 1 owner per step, and no step that begins before the one under it has printed its result. 1 order, 0 improvisation.

The last step is a table, and the table is the asset. 9 symptoms, 9 fixes: a key that is not set, a 401 from the voice provider, a 2013 quota window from the generation plan, a 401 that is a region mismatch rather than a bad key, a missing ffmpeg filter that is not missing at all because captions are drawn with the imaging library, a file-not-found that is a frame count, waxy skin that wants a re-render with biological cues rather than a retouch pass, a robotic voice that wants a lower stability figure and a higher style figure, and a quota that simply needs 5 hours to reset. Every one of those failures has already happened to someone. Written down, the next person starts with the fix instead of the fault. 393 pages and 12 countries of building taught me that the asset is never the session. It is the file the next session starts from. The probability that a machine set up twice from memory fails in the same place twice is not zero. Most people price it at zero, then debug the same afternoon again.