Scene Direction¶
The Director is an optional model pass that runs before the Writer. It reads the conversation and prepares a Scene Guidance block with guidance about mood, pacing, focus, or other details. The Writer sees the block; it is hidden from the conversation view and available in the Inspector.
You control the Director with fragments. Orb includes default fragments, and you can add, edit, reorder, or disable them.
Mood fragments¶
A mood fragment is an instruction that the Director turns on or off for a turn.
Think dynamic prompt management. For example, a Terse mood can tell the Writer
to use short sentences during a tense scene.
Each mood has:
- ID: an internal name using letters, numbers, underscores, or hyphens
- Label: the display name shown in the sidepanel
- Description: when the Director should use the mood
- Prompt text: instructions sent to the Writer while it is active
- Negative prompt: optional instructions sent to the Writer when the mood is turned off
When a cooldown makes a mood unavailable, Orb pauses it without sending its negative prompt. Negative prompts are reserved for ordinary Director-driven deactivation.
When Direction or the Agent is off, new replies have no active moods. Existing cooldowns still age, and saved state fragments still inject as configured.
Interactive fragments¶
An interactive fragment is a value the Director fills in for the current turn.
For example, a pacing fragment might produce slow burn or time skip.
Each interactive fragment has:
- ID, Label, and Description, with the same purpose as a mood fragment.
moodsandretireare reserved, because the Agent's tools already use them. - Injection label: the heading shown to the Writer, such as
Pacing: - Field type: the shape of the value
- Required: whether the Director must provide a value every turn
The Description is read by the Director. The Injection label is read by the Writer. Describe the value you want and include examples when useful.
Field types¶
| Type | Use |
|---|---|
| Single | One text value, such as Pacing: slow burn. |
| List | Several values shown as a list, such as active plot threads. |
| Feedback | A note shown to you after the reply. See Feedback Fragments. |
| State | A value or a list kept across turns on the conversation branch, such as trust, an inventory, or open threads. See State Fragments. |
| Post-processing | An editing instruction applied to the completed reply. See Post-processing Fragments. |
| Decision | A question the Judge answers before the Director runs, contributing guidance the Director plans around. See Decision Fragments. |
Macros and order¶
Fragment text supports macros. A random macro in mood prompt text rolls once per conversation. A random macro emitted in a Director value rolls on each turn. Put a macro in single backticks when you want the Director to see it as literal text.
Fragments run from top to bottom within their stage. Earlier Director values can provide context for later values, and the Writer sees them in the same order. Post-processing fragments run afterward as Editor tasks and do not enter the Director schema or Scene Guidance.
Inspector in the chat¶
Settings → Show Inspector in chat shows each reply's turn details above its text instead of one reply at a time in the Inspector panel. A Reasoning block has a tab for each pass that wrote some; the Inspector block under it holds moods, agent latency, decisions, feedback, state changes, tool calls and the injection block. Empty sections and empty blocks are left out.
A collapsed Inspector block keeps a summary line of active moods and decision outcomes. Opening or collapsing a block or section applies to every reply and to the panel, and is saved; the Reasoning block opens apart from the panel's Reasoning controls. The panel keeps the context size and next-turn reasoning controls.