Spiral

Most AI video tools assume filmmaking is linear: prompt in, video out. Real work isn’t like that. It’s revision, where every change ripples into everything downstream. Spiral is a tool that works that way from the start, and shows its reasoning as it goes so your judgment stays in the loop. It is what making thisness and Samudra Manthan by hand taught me to want.

It has had two versions. The first was one agent revising one film through a typed dependency graph, with locks, cascades and a converge step. That taught me the primitives, and then it hit the wall every single-agent tool hits: one context window cannot hold a director’s taste, a cinematographer’s eye for what a video model can actually generate, and a sound designer’s argument about whose clock the film runs on. Real crews split that up and then argue. So the current version is a council. You type one sentence. Seven departments take it through four gates and hand back a render sheet, an edit sheet and a mix sheet you follow in your own tools, and every argument along the way is on the record. Each section below is one mechanism in a line and a picture; the page it links to is the how.

Version one: a dependency graph that hit the single-agent wall

node story

→

edit here scene

→

node shot

→

node frame

cascade ↓

• auto-propagatetext-only, free
• flag for reviewyou decide
• hard-invalidateregenerate

locks style · character · motif · voice — injected into every prompt downstream, so the film stays one film

converge gathers what's unresolved, works out the consequences, walks you through settling them. Then, and only then, paid generation.

Version one: a film as a typed dependency graph (story, scene, shot, frame). An edit anywhere cascades downstream in one of three modes; locks for style, character, motif and voice are injected into every prompt below them; a converge step settles what is unresolved before anything paid is generated. Seven prototype iterations in plain HTML, CSS and JavaScript. Every one of those mechanics survives in the council, and the wall it hit is why there is one. The first design →

Four gates, from logline to delivery

gate →

Ideation locks logline · beats · spine

Pre-production locks script · bibles · shot list

Production locks every take you accepted

Post locks the cut · the mix · delivery

Producer

leads

active

active

leads

Director

leads

active

watches dailies

leads

Story

leads

leads

advises

advises

Look & Cast

sketches

leads

critiques

grades

Camera

advises

leads

critiques

advises

Sound

sets the spine

temp track

leads

leads

Unit + Edit

—

runs tests

leads

leads

leads active advises · critiques not at this gate

Four gates: ideation locks the logline, the beats and the spine; pre-production locks the script, the bibles and the shot list; production locks every take you accepted; post locks the cut, the mix and delivery. You have three jobs across all of them (choose the treatment, answer what cannot be guessed, judge the takes) and a fourth when a disagreement outlives its rounds. The gates, and how one closes →

A department is a row in a table and a five-step wake

code assemble decides what this task sees: the locks that bind it, what changed since it last read them, its open flags, its last rationale. The bundle is recorded.

→

model generate writes one part of the artifact it owns, and may FLAG, ASK or OFFER on the side.

→

model · a different prompt critique scores the proposal on three named dimensions against the locks; can force one revision before commit.

→

code resolve pins the answer to the ids it asked for, merges the part in, bumps the lock, routes flags to the Director and asks to the Producer.

→

code record appends every step, with its cause and its cost, to the log. Then the department sleeps.

A department is a row in a table: the one artifact it owns, the locks it subscribes to, the question its critic asks. Every department, at every gate, runs through the same five-step wake. The design had nine departments; the build has seven, and the two merges are one row each. Departments are data → · One harness, five roles →

One writer per artifact; the Director rules

targets your own artifact PROPOSE a new part of the one thing you own. Becomes a commit when the resolver accepts it.

targets someone else’s artifact · against a lock FLAG a concern, with a severity: note · revise · block. Read-only; it cannot change anything by itself.

targets the human, via the Producer ASK a question only you can answer, with why it matters, a best guess, and the undo cost of guessing.

targets any artifact · at most one per wake OFFER something nobody asked for, noticed while reading another department’s work, with a text preview.

One writer per artifact. A department that thinks another’s work is wrong raises a flag against a named lock; the Director rules; the owner revises. No department ever calls another. Every gap in the brief becomes a guess if it is cheap to undo, or a question for you if it is not. The four messages and the four rules →

A gate closes on a table, not a feeling

seen ↓ · current →

visionv1

spinev1

scriptv5

biblesv2

shot_listv3

audiov2

open flags

Produceractive

v1

v1

v5

v2

v3

v2

none

Directoractive

v1

v1

v5

v2

v3

v2

none

Storylead

v1

v1

–

v2

–

v2

none

Look & Castlead

v1

–

v4

–

–

–

none

Cameralead

–

v1

v4

v2

–

–

1

Soundtemp track

v1

v1

v4

–

v3

–

none

Unit + Edittests

–

–

–

v2

v3

v2

none

not converged script moved to v5; Look & Cast, Camera and Sound last read v4 and will be woken as critics. Camera has one flag open. The gate closes when every cell is green and no revise or block flag is open.

A gate closes on a table, not a feeling: no revise-or-block flag open, and every active department has evaluated the current version of every lock that binds it. This is version one’s cascade made honest: a change to the script does not propagate to Camera, it makes Camera stale, and stale departments are woken to re-read. Rounds are capped; loops end in a question, never in silence. Convergence, staleness and the round cap →

There is no database: every screen is a projection of the log

one event

id

evt_m1x…

at

10:42:18

gate · round

ideation · 3

actor

camera · critic · model · camera@v2

caused_by

evt_m1w… (artifact.committed)

type

flag.raised

target

script an artifact, never a department

against

spine@v1

payload

severity · reason · suggested

cost

1.8k in · 0.4k out · $0.011 · 6.2s

why? · walk the chain back

1. gate.opened you “a 10-second clip of high fashion in the slums of India”
2. ask.answered you shoot treatment B: the fitting
3. lock.changed director vision → v1 · spine → image_led
4. dept.woke story writing the beats · bundle recorded
5. artifact.committed story script v1 · the turn is in beat 3
6. flag.raised camera on script, against spine · revise: beat 2 needs hands
7. ruling.issued director uphold · Story keeps the hands off screen
8. artifact.committed story script v2 · resolves the flag

projections · rebuilt from the log

• board
• margins
• call sheet
• render sheet
• dailies
• ledger
• inspector

If one is ever wrong, delete it and replay. Drag the scrubber and every screen shows the project as it stood after event N.

There is no database. Every change is an appended event with an actor, a cause and a cost, and every screen is a projection rebuilt from the log. So “why did this happen?” walks a chain back to the sentence you typed, “what did it see when it decided?” is the recorded bundle, and the whole film can be scrubbed back to any event. The log is the film → · One loop, nine events →

Surprises live in offers, never in locks

sound · generator, one extra question “reading this, is there something worth offering?” noticed while: reading script v2 — beat 4 ends on a held frame with no sound line

→

offer.made · targets the audio timeline drop the score two bars early preview: the cue as text · if accepted: the last 2.5s carry only the street · cost so far: nothing

→

you · on the call sheet

try itnot now

tried → the owner incorporates it and every stale critic re-reads · passed → recorded in the dailies as offered, not taken

Every generator is asked one extra question on every wake: reading this, is there something worth offering that nobody asked for? Surprises live in proposals, never in locks (the film is offered to you, not changed under you) and the dailies keep the roads not taken. Engineering surprise →

The council writes text; you render

the council writes · text only the render sheet

• referencesone image prompt per character and environment in the bible
• shotsone video prompt each, carrying that shot’s native-sound line and the standing negative verbatim
• music · voice · SFXone prompt each for a text-to-music or voice tool

→

you render · in your own tools copy · paste · generate then report, per take, with a note

accept reject → rewrite skip

→

a rejection is an event the owner rewrites that one prompt “change what the note names; keep the same subject description, the same native sound line, the same negatives.” Every stale critic re-reads, cheaply.

the gate closes when every item on the sheet has a take accepted or skipped, no revise-or-block flag is open, and every critic has read the current sheet. The budget is the number of times you clicked accept.

The council writes text. Nothing renders inside it. You render every prompt in your own tools and report accept, reject or skip per take; a rejection with a note rewrites that one prompt. Iteration in text is free; anything that costs money is a click you made. That line was drawn in version one and has not moved. The render sheet → · The board, and its three depths →

A working prototype, with no film through it yet

1. v0 · built the log and one worker planned: an events table, projections as views, one worker that wakes departments by subscription in the prototype: an append-only array in the browser, projections rebuilt on every event, the five-role harness, subscription and staleness
2. v1 · not built durable execution planned: workflows that sleep for days waiting on a human, activities with retries and heartbeats in the prototype: a wake is a chain of fetches in a tab; close the tab mid-round and the round is lost, though the log is not
3. v2 · built the board planned: board, margins, call sheet, ledger over server-sent events; inspector last in the prototype: all six surfaces, including the inspector, the scrubber and export / import of the whole diary
4. v3 · not built telemetry and provenance planned: OpenTelemetry spans; a manifest on every rendered asset; signed credentials at export in the prototype: frames are rendered by hand outside the council, so there is nothing yet to attach a manifest to
5. v4 · built offers and dailies planned: the extra question on every wake, a surprise budget, the recorder’s digest in the prototype: one offer per department per wake; dailies written from the log at every gate close, with the roads not taken

A working prototype, in a browser, one model playing every role, no film delivered through it yet. The log, the board, offers and dailies are real; durable execution and provenance are not. Two of the design’s bets are still untested, and the question version one left open, one user to many, is still open. Design versus build → · The design document →

Go deeper

• The first design: version one, the dependency graph, the cascades and the locks, and the wall it hit.
• Departments are data: seven rows, what each owns, and why nine became seven.
• One harness, five roles: assemble, generate, critique, resolve, record; who sees what.
• Four messages, four rules: propose, flag, ask, offer; single writer; the Director rules; guess or ask.
• Gates and convergence: the stage table, the staleness table, the round cap.
• The log is the film: events, projections, the why chain, time travel, the ledger, the dailies.
• Engineering surprise: offers, previews, the roads not taken.
• The render sheet: production and post, by hand.
• The board: six surfaces, three depths.
• One loop, nine events: a script rewrite, settled without an orchestrator.
• Against the plan: what the prototype built, skipped, and has not yet tested.
• The design document: the September 2026 proposal the council was built from.
• The prototype: run it.