Directing the AI performer#
Week 8 — SLOP4555
Every generation is a take, not a result#
Rehearsal implies you expected some of them to fail. Treating each generation as a surprise is the mistake this week corrects.
If nothing is ever rejected, nothing is actually being directed.
Diagnose against objective and tactic#
Mamet’s frame: a shot argues something, through a tactic, toward an objective. Use it as a rubric — does the take serve that argument?
| Area to investigate | What to check |
|---|---|
| The note | Playable circumstance, visible timing, conflicting instructions |
| The reference | Actual opening pose, geography, bible and supplied inputs |
| Model/workflow | Supported controls and variability under a clear setup |
Observation → possible cause with uncertainty → next check. A mismatch alone does not prove which cause was responsible.
A real rejected take#
Week 7 B: real sora-2 output, same medium starting frame.
- Intent: reach, stop short and hold without touching.
- Observation: around 3–4 seconds, fingers reach the phone edge and stay.
- Decision: reject; the stopping gap is missing.
- Diagnosis: motion failed to preserve the no-contact boundary; inspect note and reference before assigning a cause.
- Later direction: preserve a visible gap and hold hovering.
Sora A was blocked with no footage. Later Jimeng A reaches around 4.5s; Jimeng B around 2.5–3s, then holds hovering. B still misses immediate response. Watch the candidates in Week 7. Platform and prompts changed together: no causal repair or completed selection is claimed. Audio/review gaps remain.
Actionable direction, not “more/better”#
“Make it better” has nowhere to land. A note aimed at a named failure — “hold the stillness half a second longer before the reach begins” — does.
Take selection is a discipline#
Choosing the best of several takes and being able to say why — against the beat’s original intent, not against “which one looked best.”
The same skill as a script supervisor’s: not taste, accountability.
This week’s task#
Bring one performance beat that hasn’t worked cleanly yet. Generate three takes. For each rejected one, record the mismatch, possible cause and next check. In review: observe first, reveal direction/reference, then discuss causes. Select your take and justify it against the beat’s objective and tactic.
What you leave with#
Three takes, a diagnosis for every rejection, and a selection you can defend.
This diagnosis log is what a director’s shot log looks like — keep the habit for every beat from here on.