← All work

Creative workflows · Local-first

Storyspan 源映

Carrying a director’s creative intent into a deliverable screenplay.

Sole designer & builder — product architecture and creative workflows

Screenplay main chain wired through; first-release work remains. Film output is paused; no finished film yet.

CREATIVE INTENT → DELIVERABLE

  1. 01

    Knowledge

    Sources · extraction

    Traceable references
  2. 02

    Screenplay

    Scenes · approval

    Fountain export
  3. 03

    Prompt package

    Freeze · gap report

    External handoff

The screenplay is the product. Film output is paused.

Workflow illustration · not an application screenshot

01The project

Storyspan distills source material into traceable knowledge assets, develops creative intent into a deliverable screenplay, and compiles it into a frozen prompt package for handoff to an external executor. The screenplay is the product, and the finished film is the creator's own internal output.

An independently designed and built creative-technology system, local-first, formerly Director Intent OS. Its product architecture — Charter v7, four modules — was frozen on 2026-08-27, and the product was named Storyspan (源映) on 2026-09-04.

What this case provesStarting from a real industry pain point and turning the director's tacit judgment into an auditable, deliverable product workflow

02The problem

A film is a chain of interdependent creative decisions: source interpretation, story, screenplay, staging, performance, and continuity. A single model call cannot retain the director's judgment, continuity across the film, or the revision history; as the work goes on, the creator can no longer say which version was accepted, what changed, or why. What Storyspan sets out to solve is making creative intent traceable, auditable, and deliverable across the whole chain.

03Key decisions

Storyspan runs four modules on one governed chain. Knowledge extraction works inside declared, versioned frameworks, and extracted knowledge counts only after human signoff. Creation treats the story trunk as the single source with three peer entries — an original concept, a rights-cleared adaptation, or knowledge-assisted assembly — carrying adaptation contracts and production versions, with story bibles, per-scene revision approval, three checks, lock and release records, and Fountain export; that main chain is wired end to end. Generation preparation and compilation turn an approved screenplay into storyboard and reference bindings, then into a model-neutral generation contract, compiled under one target model's official format specification; when something cannot be expressed, compilation refuses and reports the capability gap, and what compiles is a frozen prompt package. Film Run — one external executor, one playable file, one run ledger — is paused.

04My contribution

  • Versioned knowledge extraction: extraction runs inside declared, numbered frameworks; old atoms keep their old version and are never retroactively rewritten, and knowledge counts only after human signoff
  • Story trunk and adaptation contracts: three peer entry points (original, rights-cleared adaptation, knowledge-assisted assembly); one trunk carries single-film or episodic contracts; aspect ratio, style, language, and release versions settle at the production-version layer
  • Per-scene revision approval, the three checks, lock and release records, and Fountain export
More contributions
  • Generation preparation and compilation: model-neutral generation contracts compiled into a frozen prompt package under the target model's official format specification, capability gaps reported explicitly, and formats without an executor marked manual-export-only
  • A four-module product charter voted on item by item, plus the three gates of design review, implementation review, and final acceptance

05Design in practice

Product positioning

The product of Storyspan (源映) is the screenplay: a screenplay that can be handed to someone else and accepted by them, carrying complete delivery records — which version, who approved it, when it was locked, and the file fingerprint.

The finished film is the creator's own work and stays an internal output. The film-output stage therefore keeps only its thinnest layer: hand the frozen prompt package to one external executor, get back a playable file, and keep one run ledger. The prompt package itself is an internal artifact — viewable, freezable, exportable, and hand-carryable to any external tool — not priced separately and making no picture-quality promises. This division keeps the product's promises tied to what it can deliver for acceptance.

Knowledge with a version number

Extraction only runs inside a declared framework, and every framework carries a version number. When a framework is revised, old extracted atoms keep their old version; nothing already accepted is rewritten in hindsight. That rule keeps knowledge auditable long after the fact: any statement in the system can be traced back to the framework version and the source material it rests on.

Six knowledge angles are live — writing, scene, directing, props, continuity, and generation constraints. Coverage of each angle's own frameworks is an on-demand breadth item, not a completed angle-by-framework matrix.

A real drill put numbers on the boundary: 269 paragraphs of a textbook produced 19 candidates, each reviewed one by one under the owner's delegation — 9 adopted, 10 rejected. Adopted items were rewritten by hand before promotion. When a later review finds an extraction wrong, the error can be marked and the works in progress that depend on it get alerted.

The screenplay as the product

The story trunk is the single source, entered three ways with equal standing: an original concept, an adaptation with clear rights, or knowledge-assisted assembly from the extraction module. One trunk can carry several adaptation contracts — a single film or an episodic series — and choices like aspect ratio, visual style, language, and release versions live at the production-version layer under each contract.

Character, relationship, and scene bibles take shape from actual usage; a short film gets a bible that is narrower, not shallower. Scene revisions carry per-item approval as recorded semantics, and the three checks — change impact, continuity, script doctor — run at the outline, scene, and lock gates on an acknowledged baseline with incremental projections. A blocking finding is advisory and does not stop a lock.

After a lock, the system issues a formal release record and the screenplay exports to Fountain. Start the project, write the scenes, run the checks, approve the lock, export — that main chain is wired end to end.

Compile, do not translate

This module is called generation preparation and compilation because it does not paraphrase a screenplay into prose for a model. It compiles: an approved screenplay becomes storyboard and reference bindings, then a model-neutral generation contract, and finally an artifact written in one target model's official format specification. The frozen prompt package is an internal artifact — the handoff format for an external executor — not a product on its own.

When a screenplay asks for something the compile target cannot express, compilation refuses and reports the capability gap explicitly instead of silently dropping the request. That fail-closed behavior has been delivered.

Quality conclusions use a three-level vocabulary: format-legal, server-accepted, and generation-quality-verified. The system presents the conclusion and never adjudicates it for the creator; until a real executor is connected, it can give at most the first level.

Every frozen package pins its target model, format specification, and compiler version. Any change upstream — screenplay, storyboards, reference assets, or compiler — marks that package stale instead of silently rewriting it.

Compile targets are many; execution targets are one. A format with no attached executor is marked manual-export-only, so the boundary of what can actually run stays visible.

How it is built

Architecture first, then code. The product architecture is an 89-item charter, each item voted on line by line by the creator and the design side, then handed to independent review for repeated challenge until it froze without dispute; every resolution keeps its reason on record, and every later feature can be traced back to the resolution it rests on.

Delivery proceeds by feature packages: the ones that settle the data shape go first, in order, and extensibility features are pulled on demand. Every package passes three gates — design review, implementation review, and final acceptance — before it merges, with no exception among the 48 packages.

The division of labour is explicit: design, independent review, and implementation each do their own part, and the creator holds final adjudication. All records are append-only and never overwritten, so how this product was built, step by step, is itself auditable.

06Evaluation

Two sets of real numbers. Knowledge extraction ran a live drill on a textbook (2026-08-28): 269 paragraphs of text produced 19 candidates, and a line-by-line review adopted 9 and rejected 10, validating the gate that the machine only proposes candidates and human signoff makes them count. By 2026-09-04, the product had delivered 48 feature packages under the four-module charter, each one passing design review, implementation review, and final acceptance before merge.

07Outcome and current stage

Knowledge extraction is usable. The screenplay main chain is wired through from project start to export: create the project, write the scenes, run the checks, approve the lock, and export Fountain; the per-scene signoff interfaces, background scheduling for the checks, AI drafting context, and automatic backups will be completed before the first formal release. Generation preparation and compilation can already produce a frozen prompt package with a capability-gap report and judge whether a package has gone stale. The film-output stage is paused, and the system has not produced a finished film.

Technical details & evidence

Role · Sole designer and builder — product architecture and scope adjudication, four-module design, the screenplay product main chain, generation preparation and compilation, and a multi-seat AI collaboration development process

  • Append-only records with immutable revisions: decisions and approvals accumulate as history instead of being overwritten
  • Fail-closed gates: a stage stops advancing when a required human signoff is missing or a requirement cannot be expressed
  • Multiple compile targets, one execution target: a format without an attached executor is marked manual-export-only instead of pretending to be runnable
  • Truth documents are protected by versioning, plus a feature-reachability check: a finished feature must be reachable by someone in the interface
  • Python
  • FastAPI
  • SQLite
  • SQLAlchemy / Alembic
  • Jinja2
  • FFmpeg / ffprobe
  • Local multimodal AI