Choose the behavior mode before rewriting your prompt
If the AI moves past your turn, check the behavior setting.
- Take-Turns
You want to answer for your character
The AI detects your turn and pauses. It can still introduce new characters.
- Spectate
You want to watch the whole cast
The AI plays everyone, including your character. Choose this when you want a scene to unfold without taking each turn.
- NPC Only
You want to keep the existing cast
It uses defined or previously introduced characters. It does not detect your turn; output or interaction limits stop it.
These are documented behaviors, not a reliability test. Official behavior settings.
For a two-character negotiation, I would use Take-Turns to preserve my response.
Correct the session’s scenario, not only the original
V2 gives each session its own scenario copy. Editing that copy does not alter the original, and changes to the original do not automatically update an ongoing session. If an old character detail keeps returning, check which copy you edited. V2’s scenario-copy rules.
Open the session editor in the desktop sidebar or mobile Book icon. Hidden reaches the AI; Excluded does not. Screen visibility does not establish context inclusion. Editor access and interaction flags.
Put the most important demonstration first
DreamGen distinguishes the Opening from Examples. Opening text can eventually leave context as a session grows. Examples serve as continuing reference material, but not all necessarily fit; the first ones receive priority. How openings and examples differ.
I would use one short example that demonstrates the character’s essential rule before adding decorative backstory. A concise contradiction-free reference gives you a clearer starting point for diagnosis than several pages of overlapping instructions. This is an editing recommendation, not a measured recall improvement.
Proposed check · Not performedFind the source of a recurring canon error
Use a harmless fictional scene with two adult characters. For example: Mara owns a brass compass; Leon has never seen it.
- Record the model, provider, behavior mode and context setting. Keep them fixed for the first comparison.
- Inspect the active session’s character and plot notes. Remove contradictory ownership details, and keep a separate copy of your intended canon.
- Check whether the relevant interaction is Excluded. Correct the supplied context before asking for a short scene that depends on who knows about the compass.
- Compare several continuations after unrelated conversation. Record contradictions and any manual reminders rather than keeping only the best reply.
An immediate correct answer after a correction shows that the model can use recent supplied information. It does not prove durable memory. A later error could reflect omitted context, conflicting notes or model behavior; this exercise does not isolate all three by itself.
Use the Writing Assistant for the current scene’s problem
Story mode offers editable text blocks and + Instruction for directing the next passage. Its Writing Assistant can see the whole scenario and the most recent story text. That documented scope is narrower than reading an entire long manuscript. Story controls and assistant scope.
For a late-chapter reveal, I would put the necessary earlier clue in the scenario or supply a short recap before asking the assistant to assess it. Separate two questions: “Does this scene work?” and “Is this consistent with chapter one?” The second requires evidence from chapter one.
If your priority is developing prose around a managed story reference, compare NovelAI’s writing workflow. If choosing a model connection is central to your setup, read the SillyTavern review. Neither comparison establishes that one model writes better.
Advanced and Pro list the same monthly credits
Daily credits apply only after monthly credits run out and do not accumulate.
Context is model-specific: the FAQ lists Advanced GLM 5.2 at 8k and Pro Gemma 4 31B at 60k, despite 15k/30k card headlines. Check the model and provider before upgrading.
Broad “unlimited” card claims need the FAQ: Free/Starter spend credits; Advanced exempts selected smaller writing models; Pro exempts first-party writing models. Images cost credits, fair-use limits apply, and provider matters. Current plan cards and expanded FAQs.
A first-party label is not a promise that data never leaves DreamGen
The March 23 privacy policy says model-improvement consent is off by default. Withdrawing consent stops new content being used for model improvement; it does not retroactively remove previously opted-in data from training datasets or model weights. It also permits vetted third-party routing under specified conditions, including demand or service disruptions. These are vendor policy statements, not a privacy audit. Training consent and AI processing policy.