Generating locations

Locations are the foundation. A useful one is not merely a pretty frame: actors can be blocked inside it, the light has one logic, the camera can see depth, and a later view can repeat the same geography. Slow down here, because every shot inherits these decisions.

Turn the rough idea into a production contract

There's no universal sentence formula. Start messy — "a big three-storey farmer's house in the middle of nowhere, red jeep out front, hot summer day" — then make the decisions that the rough idea leaves open. Claude can assemble the wording, but it should not silently direct the location.

Concept pass — the messy description translated once. It proves the broad subject, weather, and red vehicle, but it does not yet prove blocking or reverse-angle continuity.

One note later — the visible changes land: house in the right third, fenced pens on the left, warm low sunset light. That is evidence for composition and light, not proof that an unseen reverse will preserve the drive, pens, or house geometry.

Ask Claude to rewrite the full prompt and return a decision log. Compare that log with the five locks above; if a lock disappears, restore it before generating. Conversation makes revision fast, but the evidence — not Claude's confidence — decides whether the location is usable.

The rules that keep a location believable

A believable location is a chain of causal decisions. The lab gives you one scene brief: choose the geography, blocking anchor, light logic, materials, and depth/continuity lock that make it shootable. Then compare an approved reference with a failed counterexample. These are diagnostic generations, not a controlled scientific A/B; surrounding drift is itself evidence of what the prompt failed to hold. The final prompt and approval rule use only the decisions you proved, so every clause has a visible source.

Two failures are easy to skip past. An oversharpened, oily surface has no real texture to hold on to, so it swims and smears the instant anything moves across it. And atmospheric haze is what tells the eye a background is far away — kill it and everything, near and far, reads at the same sharpness: flat, and immediately readable as slop.

These two references pass the light, tonality, texture, and depth checks visible in one still. Neither image alone proves a blocking route or reverse continuity:

Misty forest — smooth tonal steps from shadow to sky, nothing crushed, fog as the one soft light source. Passes.

Boulevard at sunset — one sun, shadows all agree, neon reads as real signage. Passes.

Shoot at 3/4, then test the reverse

A 3/4 view exposes side geometry and separates foreground, midground, and background. That gives later character placement visible floor and distance cues. A reverse is a separate generation and a continuity test: keep the anchor, openings, light side, materials, and palette locked, then compare the two frames before calling them the same location.

Shoot head-on and the location can flatten into a backdrop. The hidden side geometry remains unconstrained, so asking for an angle later gives the model more room to invent doors, walls, and object positions.

Give the frame an anchor

A location needs an anchor — a clear object you attach blocking to: a big sofa, a front door, a specific street banner. Replace “the character stands on the left” with “the character stands between the sofa's hall-side arm and the window.” The relationship is testable in a later frame; the vague screen direction is not.

When one master wide is enough for b-roll

When no cut depends on exact off-frame geography — some b-roll, montage, and action coverage — one wide plate can be enough. Mediums and close-ups may extend what the plate never showed, so use this shortcut only when that invention cannot break blocking or continuity.

Every still came from that single aerial wide. They pass as a montage because the track, curbs, barriers, season, and palette hold; they would not prove an exact reverse or a repeatable turn.

Lock in the rules