Chapter 0 — Director and agent
Status: DRAFT 2 (2026-09-21), rewritten to DESIGN v6 §5 after the director and professor reviews. Draft 1 opened with nine rules and a ten-row table; this draft opens with what went wrong and keeps three rules. The full rule table with each rule's "why" is in Appendix E.
0.0 Objective
Given three agent reports, you can mark every verdict the evidence does not support. Given a card's goal line and two takes, you can pick one and say why in one sentence.
0.1 What you are deciding
You are about to make a film with an AI agent doing most of the typing. Before a single image exists you decide one thing: who decides what, and how the two of you talk. Get this wrong and the agent will generate confidently, spend your money quietly, and tell you a shot "passed" when nobody watched it. Get it right and the agent becomes what it was for us: a tireless operator that measures everything, spends nothing without a fence, and reports in a form you can act on in one minute.
This chapter produces one document, your working agreement, and installs three habits: the fence, the ledger row, the report. It also gives you the one method a first-time director needs most and has least: how to choose between two takes when you have no taste yet.
0.2 How we did it
Three rules, one table
HALFWAY had one human in the Director's chair and two AI systems as operators. The split that mattered fits in one line from our process document: no clip enters the assembly without a Director-selected take. Not "the one that passed"; the one that was chosen, with one line of why in the log. If nobody chose, the shot is not done.
The three rules the agent worked under from the first hour (the other six arrive in the chapters where they first bite; all nine with sources in Appendix E):
| Rule | Why it exists |
|---|---|
| The Director approves every spend. One job per approved decision; the agent reports after every job and every read. | An agent that can spend unsupervised will, in good faith, buy a fourth take of a shot the Director would have cut. |
| Ledger every job, every read, every decision. One row each, appended, never edited. | Without it the balance is a guess. Ours was still 14 credits short of the platform's own export at the end (Appendix A §A.2): two video jobs and a few images never got a row. The platform's number is the one we publish; the ledger's job was to make the gap findable. |
| Numbers only after the read. A price comes from the button, a balance from the account, a frame count from the verification file. No source, no number. | Numbers are what the Director decides on. Case 0-A. |
The report
The Director asked for one thing in the takeover brief: report first, then keep working. The format that survived the whole production has three lines and nothing else:
WHAT I READ — the files, strips, measurements actually looked at, with paths
REMAINING ISSUES — numbered, each with the evidence, none with a verdict the evidence does not support
NEXT ACTION — one concrete step, its cost if any, waiting on which approval
One sentence from a real one (production/phase3/post-submission/CLAUDE-TAKEOVER-2026-09-15.md): "Strips are frame samples, not playback; audio was measured, not heard." Every measurement in that report names its image and its numbers; every judgment is left to the Director.
Choosing a take
The agent can tell you a take's frame count, its loudness, whether the mouth moved inside the line's window. It cannot tell you which take is the take. Neither could we, at first. The protocol we ended with:
- Read the card's goal line once (what this shot is for; Chapter 5 writes it).
- Play A, then B, full speed, sound on. Do not pause.
- Play both again, silent.
- Write one sentence: which, and why, judged against the goal line, not against the prompt. "B: her shoulders drop before the smile, A smiles first" is a sentence. "B looks better" is not.
- Only now read the agent's measurements. If they contradict your sentence, watch again; keep both notes.
Step 5's order is the point. Impression first, numbers second. A contact sheet cannot show you that a hand moves at the right speed; a waveform cannot tell you a line was said with intention. Your eye is the instrument the measurements exist to serve, and it is the one that gets trained by use.
0.3 Case 0-A — a number written before it was read
Two rows from our ledger (production/phase3/ledger.csv), 25 seconds apart:
CUT21,edit,ffmpeg,720p24fps,2026-09-16T18:23:30Z,technical-pass-not-final,0,"cut21-…: CUT20 + two small changes. … 4453 frames 185.54s decode exit0."
CUT21,correction,ledger,frame count,2026-09-16T18:23:55Z,corrected,0,"Previous CUT21 row wrote 4453 frames 185.54s before reading verification.json; actual 4465 frames 186.041667s. Protocol miss: number written before the read."
- What I saw: the first row's frame count and duration do not match the build's own verification file, read 25 seconds later.
- Why I think so: the agent wrote the row from the plan it expected to render, not from the file the render produced. The numbers are the plan's numbers.
- What I changed: a correction row, never an edit. From then on, the row is written after
verification.jsonis opened, and quotes it. - Would this happen on your tool? Yes. Any agent, any tool. It is the default behaviour of something that can predict the answer.
The 12-frame gap is half a second. The Director was placing subtitles to the frame that day.
0.4 Do it with your tools
The agent can be any coding agent that runs shell commands, reads files and drives a browser or an API. Three properties matter; the rest is in Appendix B.
| Property | Why | If yours lacks it |
|---|---|---|
| Appends to files without rewriting them | the ledger and fences must be append-only | keep the ledger in a spreadsheet the agent cannot edit |
| Can be told "never" and keep it across sessions | the working agreement must survive a restart | put the agreement in the file the agent reads at start (for Claude Code: CLAUDE.md) |
| A second model or person to argue with | our story grader, audio plan and this book were all revised by an adversarial pass | your second reader with the checklist |
0.5 Your turn
Time cap: one evening. Spend: none.
- Write your working agreement (template below; full set in Appendix C). Fill the decision line with your own names; keep "which take is the take" on the human side.
- Create the ledger with the six columns and write row 1 by hand: today's date,
decision, "working agreement written". - Drill the report. Give the agent any short video (a phone clip is fine) and ask for a report in the three-line format on what it can measure: duration, frame count, loudness, a 6 fps strip. Reject any line that contains a judgment the measurement cannot support ("the acting is flat", "the sync is fine"). Repeat until the report is measurement only.
- Drill the marking. Below are three short reports. Two are seeded with unsupported verdicts; one is clean. Mark every sentence that claims more than its evidence. Answers are at the end of the chapter; do not look first.
- Drill the take choice. Take two phone clips of the same thing (someone picking up a cup, twice). Write a one-line goal ("she hesitates before she lifts it"). Run the five-step protocol. Your sentence must name something visible.
- Drill the fence. Ask the agent to write a fence for an imaginary 2-credit image job, with the expected price and the check it will run afterwards. Do not run the job.
Templates
Working agreement:
DIRECTOR: <name> AGENT(S): <names> SECOND READER: <name> FILM: <title>
Decisions only the Director makes: story, casting, which take, cut/rewrite, lock, delivery.
Spend rule: every paid job needs a fence and a "go" from the Director in chat. One job per go.
Ledger rule: one row per job, read, decision, fence, correction; append only; numbers only after the read.
Report format: WHAT I READ / REMAINING ISSUES / NEXT ACTION.
Take rule: the Director chooses with the five-step protocol; the agent's measurements are read after the sentence is written.
Clicks the agent hands back to the Director: <binding submissions, unpriced buttons, anything irreversible>.
Ledger row (csv):
timestamp, kind (job|read|decision|fence|correction), subject, detail, price_or_credits, status
Fence:
FENCE — <job name>, <date>
Purpose: <the beat this serves>
Inputs: <references by file name; none from outside the project>
Settings: <model, resolution, duration, sound on/off, count>
Expected price: <read at the button before the click>
Check afterwards: <the strip, sheet or transcript that decides pass/fail; who judges>
The three reports for drill 4
Report 1
WHAT I READ — clip_a.mp4: ffprobe 24 fps, 121 frames, 5.04 s; ebur128 −23.1 LUFS; strip clip_a-6fps.png
REMAINING ISSUES — 1. Mouth open in frames 3–4 of the strip (0.5–0.67 s), before the line's window at 0.9 s. 2. The performance is convincing.
NEXT ACTION — Director playback of 0–1.5 s to rule on the early mouth.
Report 2
WHAT I READ — clip_b.mp4: ffprobe 24 fps, 120 frames, 5.00 s; whisper transcript "我必须走了"; no strip made
REMAINING ISSUES — 1. Transcript matches the line. 2. Lip-sync confirmed.
NEXT ACTION — none; take passes.
Report 3
WHAT I READ — clip_c.mp4: ffprobe 24 fps, 120 frames, 5.00 s; strip clip_c-12fps.png; ebur128 −31.4 LUFS
REMAINING ISSUES — 1. Level is 8 LU under the other takes (−23); the line window 1.2–2.6 s measures −33.9 LUFS. 2. Strip frames 14–18 (1.17–1.5 s): head turns left; the card says right. Judgment on whether the turn reads is the Director's.
NEXT ACTION — Director playback of 1.0–3.0 s; if kept, +8 dB on the line window only, fenced as an audio job at 0 credits.
0.6 Gate
Each line needs valid evidence per DESIGN §5; outcomes are pass / fail / unverified, and unverified blocks.
- The working agreement exists, names the Director's non-delegable decisions and the second reader, and is loaded by the agent at session start (evidence: the file, and the agent quoting it unprompted in a new session).
- The ledger exists with the six columns and at least one row (evidence: the file).
- The agent produced one report in the three-line format with no unsupported judgment (evidence: the report text; the judgment "unsupported" is yours, by name).
- Performance task: you marked the three reports and matched the answer key on every seeded sentence (evidence: your marked copy).
- You ran the take protocol once and your sentence names something visible in the clip (evidence: the sentence and the two clips; self-attested that you watched at full speed).
- The agent produced one fence with an expected price and a named after-check (evidence: the fence file).
- The agreement's last line lists the clicks the agent hands back to you (evidence: the file).
Pass example from HALFWAY: the takeover report of 2026-09-15, measurements with paths, judgments deferred. Fail example: the first CUT21 ledger row above, a number without a read behind it.
Answer key for drill 4
- Report 1: sentence 2 of REMAINING ISSUES ("The performance is convincing") is a judgment no listed measurement supports. Everything else is measurement or a deferral.
- Report 2: "Lip-sync confirmed" is unsupported: a transcript proves the words were said, not that the mouth formed them, and no strip was made. "none; take passes" is a verdict the agent may not give. Two seeded sentences.
- Report 3: clean. Every number has a source; the one judgment is named as the Director's.