Prompt composer and preview (#48) #100

Merged

crowmaki merged 2 commits from issue-48-prompt-composer into main

2026-10-08 21:09:07 +00:00

crowmaki commented

2026-10-08 20:59:17 +00:00

Owner

Closes #48.

The app writes H3's six-field prompt for every Shot from its current inputs and its filled Slot lists, and shows it read-only on the Shot (User Stories 93–98 and 102–103, decisions in #9). The override and reconcile (#49) and the job builder (#51) read shot_prompt.

What changed

Composer (prompt.rs). load gathers a Shot's inputs, and compose is a pure function of those inputs plus the filled ShotSlots. It writes subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape and non_diegetic_music in this order. An empty field is N/A, and music is always N/A. There is exactly one [Shot 1] and no id= header.
Tags.
- <Subject N>: numbered by first appearance in the action text (mentions and dialogue speakers), then the Setup's other subjects, then anyone a Board symbol is about.
- <Prop N>: numbered in first-mention order, the same order the slot budget attaches Props.
- <Environment>: the Scene's Location.
- <Picture N> / <Audio N>: taken from the filled Slot lists, so an excluded or over-budget reference is never named.
- Character names written into free text (framing note, identity fields, descriptions) become tags. A Character the Shot does not number reads as "someone". Prop names in descriptions stay as written.
subject_definitions.
- Identity fields: the Look's overrides, then the Character's values.
- Wardrobe: the Look's default wardrobe, or the Wardrobe item's text.
- Positional Atlas phrasing, with a row name only when the Atlas holds more than one Look. The views are listed per Parity: "face close-up, front and side views".
- Partial Preservation names the mannequin picture.
- Heard-only speakers are defined as heard but not seen. A heard-only speaker is a Character outside the Setup whose every Line in the Shot is O.S./O.C. or V.O.
- A Voice reference adds "uses as 's speaking voice timbre reference". This repeats the tag instead of #9's "their", because the ticket rules out pronouns.
- <Environment> is the description, then the lighting note (the set note, else the Lighting's phrase), then "shown in". <Prop N> is the description, then "shown in".
retention_analysis. One entry per sent image, in wired order, per the spec table:
- Keyframe first frame and last frame. The last frame time is the Shot's length on the frame grid.
- An Atlas row, full or partial.
- The mannequin's attribute_transfer.
- The location angle.
- A Board weak_reference: tiled or one picture per Panel. A clean-sketch Board drops the arrows clause.
- A Prop.
Then newly_generated for every seen subject whose Atlas row is not sent, then one entry per Voice reference.
detailed_description. The framing note, then the Notation read-backs as sentences in panel order (no "storyboard" or "panel" wording, and freehand arrows add nothing), then the action text. Dialogue chips expand as follows:
- Each chip becomes <Subject N> says: <d>[Language] <tag> text</d>, with parentheticals as inline tags.
- O.S./O.C. becomes "says from off-screen:", and V.O. becomes "says in voice-over:".
- Dual dialogue is announced once: "… speak at the same time."
- A SUBTITLE Line uses its own language and spoken text.
- Character-naming parentheticals are already moved out of the quote by the action text (#46).
overall_soundscape. The Location's Ambience plus the Scene's Ambience addition, written as one sentence.
Schema v20.
- project.spoken_language: defaults to English, and a blank value is refused.
- scene.ambience_addition.
Both have undoable setters (set_spoken_language, set_scene_ambience) and are on ProjectView/SceneView and their DTOs. CONTEXT.md now defines Spoken language and notes the Scene addition under Ambience.
Adapter. shot_prompt, set_spoken_language and set_scene_ambience are new commands, and the bindings are regenerated.
UI.
- A collapsible Prompt panel on each Shot shows the composed text read-only. It loads when opened and follows changed events while open.
- Scenes view: an Ambience addition field.
- Project properties: a Spoken language field.

Tests

tests/prompt_composer.rs (11 tests, all through ProjectSession):
- Exact full-prompt text for a bare Shot and for a fully referenced Shot (Atlas, location angle, Board, Prop, both Keyframes, two voices).
- Multi-row Atlas phrasing.
- Wardrobe partial Preservation and attribute transfer, including the mannequin sent without its Atlas.
- Slot renumbering when items are excluded.
- Board read-backs and the tiled, per-Panel and clean variants.
- V.O., dual dialogue, a moved parenthetical, the SUBTITLE language and the Project language.
- The summary fallback.
- The undoable Scene Ambience addition.
- "someone" for an unnumbered Character.
The adapter smoke test covers the three new commands.
notation.rs's schema rewind drops the two new columns.
Results: cargo test --workspace --all-features, bun test src (123) and svelte-check all pass. Clippy shows only the existing type_complexity warning in api.rs.

Notes for review

I did not open the preview in the running app.

Closes #48. The app writes H3's six-field prompt for every Shot from its current inputs and its filled Slot lists, and shows it read-only on the Shot (User Stories 93–98 and 102–103, decisions in #9). The override and reconcile (#49) and the job builder (#51) read `shot_prompt`. ## What changed - **Composer** (`prompt.rs`). `load` gathers a Shot's inputs, and `compose` is a pure function of those inputs plus the filled `ShotSlots`. It writes `subject_definitions`, `summary`, `retention_analysis`, `detailed_description`, `overall_soundscape` and `non_diegetic_music` in this order. An empty field is `N/A`, and music is always `N/A`. There is exactly one `[Shot 1]` and no `id=` header. - **Tags.** - `<Subject N>`: numbered by first appearance in the action text (mentions and dialogue speakers), then the Setup's other subjects, then anyone a Board symbol is about. - `<Prop N>`: numbered in first-mention order, the same order the slot budget attaches Props. - `<Environment>`: the Scene's Location. - `<Picture N>` / `<Audio N>`: taken from the filled Slot lists, so an excluded or over-budget reference is never named. - Character names written into free text (framing note, identity fields, descriptions) become tags. A Character the Shot does not number reads as "someone". Prop names in descriptions stay as written. - **subject_definitions.** - Identity fields: the Look's overrides, then the Character's values. - Wardrobe: the Look's default wardrobe, or the Wardrobe item's text. - Positional Atlas phrasing, with a row name only when the Atlas holds more than one Look. The views are listed per Parity: "face close-up, front and side views". - Partial Preservation names the mannequin picture. - Heard-only speakers are defined as heard but not seen. A heard-only speaker is a Character outside the Setup whose every Line in the Shot is O.S./O.C. or V.O. - A Voice reference adds "uses <Audio N> as <Subject N>'s speaking voice timbre reference". This repeats the tag instead of #9's "their", because the ticket rules out pronouns. - `<Environment>` is the description, then the lighting note (the set note, else the Lighting's phrase), then "shown in". `<Prop N>` is the description, then "shown in". - **retention_analysis.** One entry per sent image, in wired order, per the spec table: - Keyframe first frame and last frame. The last frame time is the Shot's length on the frame grid. - An Atlas row, full or partial. - The mannequin's `attribute_transfer`. - The location angle. - A Board `weak_reference`: tiled or one picture per Panel. A clean-sketch Board drops the arrows clause. - A Prop. Then `newly_generated` for every seen subject whose Atlas row is not sent, then one entry per Voice reference. - **detailed_description.** The framing note, then the Notation read-backs as sentences in panel order (no "storyboard" or "panel" wording, and freehand arrows add nothing), then the action text. Dialogue chips expand as follows: - Each chip becomes `<Subject N> says: <d>[Language] <tag> text</d>`, with parentheticals as inline tags. - O.S./O.C. becomes "says from off-screen:", and V.O. becomes "says in voice-over:". - Dual dialogue is announced once: "… speak at the same time." - A SUBTITLE Line uses its own language and spoken text. - Character-naming parentheticals are already moved out of the quote by the action text (#46). - **overall_soundscape.** The Location's Ambience plus the Scene's Ambience addition, written as one sentence. - **Schema v20.** - `project.spoken_language`: defaults to English, and a blank value is refused. - `scene.ambience_addition`. Both have undoable setters (`set_spoken_language`, `set_scene_ambience`) and are on `ProjectView`/`SceneView` and their DTOs. CONTEXT.md now defines Spoken language and notes the Scene addition under Ambience. - **Adapter.** `shot_prompt`, `set_spoken_language` and `set_scene_ambience` are new commands, and the bindings are regenerated. - **UI.** - A collapsible **Prompt** panel on each Shot shows the composed text read-only. It loads when opened and follows `changed` events while open. - Scenes view: an **Ambience addition** field. - Project properties: a **Spoken language** field. ## Tests - `tests/prompt_composer.rs` (11 tests, all through `ProjectSession`): - Exact full-prompt text for a bare Shot and for a fully referenced Shot (Atlas, location angle, Board, Prop, both Keyframes, two voices). - Multi-row Atlas phrasing. - Wardrobe partial Preservation and attribute transfer, including the mannequin sent without its Atlas. - Slot renumbering when items are excluded. - Board read-backs and the tiled, per-Panel and clean variants. - V.O., dual dialogue, a moved parenthetical, the SUBTITLE language and the Project language. - The summary fallback. - The undoable Scene Ambience addition. - "someone" for an unnumbered Character. - The adapter smoke test covers the three new commands. - `notation.rs`'s schema rewind drops the two new columns. - Results: `cargo test --workspace --all-features`, `bun test src` (123) and `svelte-check` all pass. Clippy shows only the existing `type_complexity` warning in `api.rs`. ## Notes for review - I did not open the preview in the running app.

crowmaki added 2 commits

2026-10-08 20:59:17 +00:00

Prompt composer and preview (#48 ) 1e043474f1

Compose H3's six-field prompt for a Shot from its current inputs and its
filled Slot lists: subject definitions with positional Atlas phrasing and
Preservation, retention entries per the spec table, Notation read-backs as
plain directions, quoted dialogue with language, delivery and inline tags,
and the Location's Ambience plus the Scene's addition as the soundscape.
Nothing is named: Characters, Props and the Location are tags, and slot
tags count only what is sent.

Adds the Project's spoken language and a Scene's Ambience addition, both
undoable, and a read-only prompt preview on each Shot.

Address review of Prompt composer (#48 ) e2d2521d76

Repeat the subject tag in the voice timbre phrase, name an Atlas row by a position enum, announce a dual dialogue speaker once, read a cast row by name, define Spoken language in the glossary, and test an outfit sent without its Atlas and an unnumbered Character in free text.

crowmaki merged commit 17f55630dc into main

2026-10-08 21:09:07 +00:00

crowmaki deleted branch issue-48-prompt-composer

2026-10-08 21:09:07 +00:00

crowmaki referenced this pull request from a commit

2026-10-08 21:09:08 +00:00

Merge pull request 'Prompt composer and preview (#48)' (#100) from issue-48-prompt-composer into main