Previs, layout and storyboarding

Expert previs conveys an idea well enough to be judged, as cheaply as possible: one fact per beat, every shot designed for what the audience must feel, timing found in the edit from stills. The camera is the audience's body; where it stands, how high and through which lens decides the emotion, so every non-neutral choice needs a stated reason. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.

REQUIRED BACKGROUND: scenario-blender-expert (execution channel, review loop, 5.2 API changes).

Toolkit: scripts/bx_previs.py (tested on 5.2.1): import sys; sys.path += ["<this skill>/scripts", "<scenario-blender-expert>/scripts"]; import bx_previs as P.

Which skill. This one: story previs and animatics judged by a director and an editor. text-image-to-blender-blockout instead when the playblast guides a render-to-real video model (Seedance): saturated color identities, no lens jump on a cut, eased moves, a hand-off sheet; an approved animatic that must become a video guide is rebuilt there. Shared: metric Z-up, roots at the feet, one camera per shot bound to a marker, to_track_quat('-Z','Y') look-at. Strokes: scenario-blender-grease-pencil. Animation after layout: scenario-blender-animation. Real rigs: scenario-blender-rigging.

Stance (the expert delta)

• Hjalti Hjalmarsson: write the beats-of-information list before any shot. One fact per line, because a viewer takes in one thing at a time. Group consecutive beats into shots; a new treatment re-orders the same facts. A missing beat means the idea never gets a fair test.
• Hjalti: theater first, then cinematic; shots cost time. A locked-off version judges the acting; the same beats told in shots ran about 2x (up to 3x) because every cut must be read (13/22/25 s vs 47/40 s). The oversimplified version is always too short and hides logistics (a character cannot cross the room in 10 frames).
• Charge recipe (Hjalti): treat every frame as a storyboard drawing. 10 frames per shot, keys CONSTANT, an A/B alternative on the next frame, the winner on the decade frame, stills exported with versioned names, timing done in a separate edit.
• Hjalti: when in doubt, eye height; lock the lens early by A/B. Two characters: split their eye heights. Low or high is a statement. Poses built for one lens do not survive a lens change.
• Dillon Gu: composition and focal length are separate decisions, and cuts can kill a mood. The audience stands where the camera is. Leaving the space removes the danger; two inanimate inserts in a row lower tension.
• Cheat for the camera, but key before you cheat (Hjalti, Charge, Renato Roldan). Resize, float, swap props per shot; in a shared timeline an unkeyed cheat breaks the earlier shots.
• Previs is the cheapest version that still conveys the idea (Hjalti, Charge). Animate only shots whose camera move is the storytelling; everything else is one still, push-ins faked on the still in the edit. Grays in Workbench, no mood lighting, mannequins before rigs. P.audit refuses a camera moving inside a still shot, keys between block starts and a non-Workbench engine.
• After each cut the eye must find one thing (Hjalti). Restage a character in front of a competing subject; keep the angle that preserves the emotion (he refused a side cut that turned Rex into a silhouette).

Establish first

Input	Changes	Default when silent
Format	every composition (Charge: lock it first)	1920x1080, 24 fps, stills at 50 %
Treatment	shot count, runtime	theater pass, then cinematic
Method	agent has no tablet	3D previs with mannequins; drawn boards only on provided art; Rigify proxies only on request (P.rigify_proxy, Rigify is off under --factory-startup: P.ensure_rigify() first)
Mood or director style	lens, height, motion per shot	neutral grammar, deviations stated per shot
Runtime budget	durations	about 1.6 to 1.9 s per beat for a cinematic cut (derived from Hjalti's slide: 25 beats, 8 shots, 40 to 47 s)
Journeys	screen direction	out left to right, back right to left (Charge)
Deliverable	outputs	shot list, contact sheet, floor plan, animatic MP4 clean plus burn-in, versioned .blend

Workflow

1. Script to beats. Asset list, then numbered beats, one fact each; mark every beat you add. GATE: P.shot_list(BEATS, SHOTS) raises on a missing, duplicated or reordered beat.
2. Set and proxies. P.setup_scene (format, Workbench, grays, Standard view transform); P.prop for real-scale set pieces; P.mannequin for characters (root at the feet with eye_height, four pivot controls as the layout subset). Grays, one pop color for the hero (Renato: red). GATE: scale reads next to a door (2.1 m) and counter (1.0 m).
3. Theater pass. One locked-off camera (cam shared, fit="all", 35 mm, mean eye height) framing the action's extremes; one block per beat group. GATE: in the P.audit rows every subject's head_in and feet_in is true at every block; theater edit timed.
4. Cinematic plan. List several openings, re-order beats, group them into shots with intent, size, lens, height, dur. P.floor_plan with the action line and journey paths before judging cameras. GATE: one side of the line; journeys follow the direction rule.
5. Blocking. For each block: sc.frame_set(f) (holds the previous pose), set only what changes, P.key_block(all_movables, f); cheats go through the same keys; then P.make_stepped. GATE: P.stepped_report == {}; the cheat check at neighboring shots.
6. Cameras. P.build_shots places each camera from size, lens and eye height (or explicit loc/aim), sets an opaque passepartout (Charge: see only the composition; P.passepartout(sc) for other cameras) and binds a marker at its first frame. P.ab_lens on key shots before posing further. GATE: P.audit(...)["problems"] == [].
7. Look. P.contact_sheet and open it. Characters merging with a dim set: P.fog_for_shot(sc, t, chars) (Hjalti: depth fog for readability, not lighting; behind the subjects, keyed to that shot), P.values_at hero vs backdrop. GATE: the visual list under Quality gates (in S6 only the pictures caught the occlusions and the cut-off wide).
8. Edit. P.render_stills, then P.animatic (image strips held dur seconds, push on stills, live shots as scene strips, sounds, burn_in). One continuous action from several angles (Pablo, Story Tools): a live shot plus cover=True shots become one scene strip soft-split per camera; retime with P.retime_cut (the marker follows), never by moving keys. GATE: runtime within budget; frames pulled from the MP4 match the cut list.
9. Iterate in a new scene pair (02_blend/02_edit) and new version numbers; nothing overwritten.

Numbers

Value	Relative to	Source
10 frames per shot block, 20 only when justified; starts on 10, 20, 30	previs timeline	Charge
Shots named in tens (sh010, sh020)	markers	Hjalti on screen
35 mm theater frame; about 60 mm for an isolating key shot	36 mm sensor	Hjalti
F-stop 5.6, clip 0.01 to 200 m on the camera rig	Hjalti's rig	Hjalti
13 / 22 / 25 s theater vs 47 / 40 s cinematic	same script	Hjalti slide 00:46:01
3 poses for an anger burst (anticipation, overshoot, settle)	one beat	Hjalti
6 fps boarding, retimed in the edit	board scene	BouncyBrain
Eye tolerance 0.15 m; drawing breaks past 70 degrees from camera	audit thresholds	[added]
Visible height = subject height x 4 / 2 / 1.15 / 0.8 / 0.6 / 0.4 / 0.25 / 0.12	ews, wide, full, mws, medium, mcu, cu, ecu	[added]
Distance = visible height x lens / sensor height (20.25 mm at 16:9)	lens swap at equal size: d2 = d1 x f2 / f1	[added]
Handheld NOISE 0.010 to 0.018 rad, 18 to 26 frame scale	camera rotation	blockout skill; peak-to-peak measured 0.66 x strength

Camera language

Choice	Reads as	Source
Eye level, medium lens, centered, static	neutral; the start for every A/B	Dillon, Hjalti
Low or tucked in a corner	hidden observer	Dillon
Wide lens from far, then a quick zoom in	being watched; someone fixing on a target	Dillon
Extreme wide lens close up	"not okay", distortion	Dillon
Long lens	informational, flat, isolates, intensifies a two-character moment	Dillon, Hjalti, Spitfire
Top view; dolly zoom (linear keys)	pressure from above; loss of control	Dillon
Off-center from behind	isolation	Dillon
Side-on movement	the most neutral screen motion	Dillon
Tilt or move ending on a detail	that detail becomes the subject	Hjalti
Faceless framing, longer shots, significant object at the edge carried by sound	routine	Dillon

Moves that feel physical (Dillon): handheld noise when a watcher or a groggy body holds the camera, a camera operator's quick zoom, the camera dropping with the character's shoulders, less motion for laziness, more for urgency; a product shot is the stillest shot. Hjalti pulls focus across a handoff (dof.focus_object). Cuts: one side of the action line; an over-the-shoulder uses the shoulder the camera already occupies, and a jump cut can mean elided routine time (Dillon); a POV sits just off the line on the camera side [added, S6].

Quality gates

Measurable: P.audit(sc, table, chars, line=(a, b), travel={name: "right"}, stepped=movables): marker switch at each shot frame and not one frame before, main subject's eye line in frame and not blocked (ray cast names the blocker, fog included), measured vs intended size, height class with an intent for every non-eye camera, one side of the action line (POV and cross exempt), screen direction per journey, block length, two inserts in a row, non-constant keys outside live blocks, opaque passepartout, cost gates; warnings lists lights. Plus the edit runtime, ffprobe streams and duration, and for live shots an edit frame rendered against the previs frame.

Visual (open each image): contact sheet (every beat visible, one focal point per shot, hero readable, cheats and occluders invisible, pose readable from this camera), floor plan (camera side, journeys), A/B sheet (isolation vs context), frames pulled from the MP4 (order, burn-ins, push, live-shot content). Comic timing and sound stay a human call (Hjalti).

Common mistakes

Mistake	Looks like	Fix
Shots sprinkled over a finished performance	generic coverage, wrong framings	design each shot from its beats (Hjalti)
Wide framing that ignores the extremes	the tallest pose or entrance clips	fit="all" and the extremes audit
POV from the true eye point	the counter or the character's own head blocks the view	cheat the POV forward and up (S6: 0.6 m forward, 1.72 m high, over the counter); the audit reports the blocker
Door or prop between camera and hero	hero hidden in the key frame	swing it away or move the camera; the contact sheet catches it
Over-the-shoulder from the far side	line crossed, hero out of frame	stay on the camera side; audit eye line
Unkeyed cheat	earlier shot changes	key_block every movable at every block
6 fps board scene as a scene strip in a 24 fps edit	plays 4x fast	export stills or retime
Every shot animated, mood lighting, full rigs	slow, the edit fights the moves, nothing judged sooner	stills plus push, live only when the move is the meaning, Workbench grays, mannequins
Coverage as separate strips or moved keys	cuts drift from the action	cover=True run, P.retime_cut
Fog world-aligned or in front	hero tinted, fog in every shot	P.fog_for_shot (behind the subjects, one shot)

Blender 5.2 notes

• bpy.ops.render.opengl fails in --background; use Workbench render.render (headless) or a camera-locked viewport in a live session.
• keyframe_insert() ignores the new-key interpolation preference (keys stay BEZIER) though "Only Insert Available" is on: call P.make_stepped.
• frame_set on a scene that is not the window scene leaves transforms stale, and a write_still render of such a scene uses the window scene's frame: P.activate, P.render_frame.
• VSE: strips, left_handle/right_handle, content_start; SceneStrip.frame_start and frame_final_* warn (removal in 6.0). Strip.split(frame, split_method='SOFT') splits from data. New scenes default to AgX: set the edit scene to the source view transform.
• marker.camera = cam binds headless; bpy.ops.marker.camera_bind needs a timeline context.
• Rigify is off under --factory-startup: bpy.ops.object.armature_basic_human_metarig_add raises "could not be found" until P.ensure_rigify().
• Selection Sets, Storypencil and Grease Pencil Tools are not bundled; use a bone collection layout. The Storyboarding app template ships an Edit scene with sync on and auto-key off in Edit (pin off in its Storyboarding workspace).
• GP v3 strokes render in a headless Workbench still.

References

• references/procedures.md: tested code for every stage (with and without the module) and the trap tests; load before writing code.
• references/expert-notes.md: each expert's principles, disagreements and deciding conditions, with timestamps; load when choosing a treatment, lens or cut.
• references/critique.md: the review rubric and the S6 findings; load at every Look gate.
• references/sources.md: the seven source videos, what each is best for, best timestamps.