Mythweaver Docs

Prompts

How a prompt is assembled from blocks, how Mythweaver picks the prompt a chat uses, and how to preview and inspect it.

A prompt is the text Mythweaver sends to the model for every reply. You build it from reusable blocks in a fixed order, and you bind it to the chats, characters or models it should apply to. You manage prompts at Prompts (/prompts), their blocks at Prompt blocks (/prompts/blocks) and custom outlets at Custom outlets (/prompts/outlets). The Blocks and Outlets links at the top of the prompts list switch between the three.

How a prompt is built

Each time a character replies, Mythweaver builds the prompt in three passes that share one template state:

  1. Every participant's and the player's description renders.
  2. Every lore entry that activated renders into its slot or outlet (see Lore).
  3. Every enabled block of the prompt renders, in the prompt's order. Template blocks are Liquid templates that read the descriptions, the lore and the rest of the chat; the chat history block becomes the transcript.

A block that renders blank is left out. The result then goes to the model in the shape its completion mode needs:

Completion mode What the model receives
Chat One message per block, in order, with the block's Role. Neighbouring blocks with the same role join into one message, and the chat history block expands into one message per turn.
Text One string: every block's text joined by blank lines, roles ignored, ending on the replying character's Name: cue.

The completion mode is a setting of the model (see Models). You never build a separate prompt for text and chat models.

Create a prompt

  1. Open Prompts at /prompts and select New prompt.
  2. Enter a Name, and optionally a Description.
  3. Leave Purpose on Roleplay for a prompt that writes chat replies.
  4. Select Save. The editor now shows Blocks in this prompt, Preview, Add a block and, for a role-play prompt, the bindings.
  5. Under Add a block, pick a block in Block and select Add, or select Create new block to write a new one. A block you create this way is added to the prompt when you save it.
  6. Drag the blocks into order and bind the prompt (see Which prompt a chat uses).

To copy an existing prompt, open it and select Duplicate. The copy is yours, keeps the same blocks in the same order with the same on and off switches, and has no bindings. To bring in a SillyTavern preset or a Mythweaver export, select Import (see Import and export).

Tip

A fresh installation comes with a site prompt named Default Roleplay, bound Global, so every chat has a working prompt from the start. Duplicate it to make your own version rather than starting from nothing.

Arrange the blocks

The Blocks in this prompt list shows each block's name, kind (template or chat history) and role. Each row has these controls:

Control What it does
Drag handle Drag the row up or down to change where the block goes in the prompt.
Eye button Turns the block off or on in this prompt only. A block that is off shows a disabled badge and is skipped.
Pencil button Opens the block in the block editor. Saving returns you to this prompt.
X button Removes the block from this prompt. The block itself is kept and stays in other prompts.

Blocks are shared: editing a block changes it in every prompt that includes it. See Blocks.

Which prompt a chat uses

A role-play prompt applies wherever it is bound. When a character replies, Mythweaver looks at every enabled role-play prompt bound to something that matches the reply, and the most specific binding wins, like CSS specificity.

Binding scopes

From most to least specific:

Scope Matches
Character in chat One character, but only in one chat.
Chat Every reply in one chat.
Character Every reply by one character, in any chat.
Model Every reply generated by one model.
Connection Every reply sent through one connection.
Global Every chat.

For example, if Default Roleplay is bound Global and your own "Noir" prompt is bound to the character Sam, Sam's replies use Noir in every chat and everyone else's use Default Roleplay. If you then bind a third prompt to one chat, every reply in that chat uses it, Sam's included, unless a Character in chat binding names Sam in that chat.

Note

A reply picks its prompt before Mythweaver chooses which connection serves the model, so a Connection binding is never matched for chat replies today. Bind to the Model instead.

To bind a prompt

  1. Open the prompt at Prompts. The bindings section is headed Bindings — the most specific match wins and lists the current bindings, or "Not bound to anything yet".
  2. Choose a Scope. The form then asks for the target that scope needs: a Connection, Model, Character or Chat, or a Character and a Chat for Character in chat.
  3. Select Bind.

To remove a binding, select Unbind on its row. Only a prompt whose Purpose is Roleplay can be bound.

To choose a chat's prompt

The Prompt field in a chat's settings (see Chat settings) sets the chat's Chat binding: picking a prompt there replaces whatever the chat was bound to, and Inherit (no chat prompt) removes it so the chat falls back to character, model and global bindings. Character in chat bindings are not affected.

Ties and ownership

Only prompts that belong to the site or to the chat's owner are candidates, plus prompts shared with the owner. A shared prompt counts only through a Chat, Character in chat or Character binding. Replies always run as the chat's owner, so the owner's prompts decide, not those of whoever pressed the button.

When several prompts are bound at the same level, the owner's own prompt wins over a site prompt, a site prompt wins over a shared one, and after that the prompt whose name sorts first wins.

If nothing matches at any level, the reply fails with "No prompt is bound for this chat."

Note

Duplicating a prompt never copies its bindings. Two prompts bound at the same level would otherwise let the name decide which one renders.

Preview

The prompt editor and the block editor each have a Preview panel that renders the prompt against a real chat without calling a model. Use it to check your templates and lore.

  1. Pick a Chat. The list holds your role-play chats; with none, the panel says "Create a chat to preview against."
  2. Pick who replies in Reply as. Default responder uses the chat's default responder, and the player is listed as "Name (player)".
  3. Select Preview.

Three buttons switch the result without running it again:

Mode What it shows
Blocks One card per rendered section, with its block name, kind, role, estimated tokens and any template errors. A block that renders nothing says "Renders nothing for this chat."
Text The whole prompt as the single string a text-completion model receives, sections joined by blank lines and ending on the replying character's Name: cue.
Injections What became of each of the chat's injections, with its position and depth, or "No injections in this chat."

The block editor previews just the block as currently typed, saved or not, and even when it is turned off. The preview remembers your chat and mode as you move between prompts and blocks, so comparing two of them takes one click each.

A preview runs the same build as a real reply, with two differences: lore timing (sticky and cooldown) does not advance, and variable writes are discarded. The panel says so ("lore is not committed and variable writes are discarded"). Token counts are estimates, since no model is chosen.

The prompt inspector

The prompt inspector shows exactly what a generated reply sent. To open it, select the Inspect prompt button in a generated message's footer (it is also in the message's menu). It describes the generation of the message's current swipe, and appears only on messages that kept a record of their prompt.

The top of the inspector summarizes the request:

Field What it shows
Prompt tokens The prompt size the provider reported.
Cached How much of the prompt the provider reused from its prompt cache, as a count and a share.
Sections The sum of the counted sections. The difference from Prompt tokens is the chat template's overhead.
Context How full the model's context limit is.
Trimmed How many old messages were left out to fit the context limit.
Completion tokens The size of the reply.
Tokenizer What counted the sections, or "none (estimated)".

A bar below shows what filled the prompt: each section, the Template overhead and the space Reserved for the reply. Six tabs then show the details:

Tab What it shows
Sections Every rendered section with its share of the prompt and any template errors.
Lore The lore entries that were Activated and those Rejected, and whether a book overflowed its budget.
Injections What became of each chat injection.
Variables The variable writes the prompt's templates made, or "No variable writes".
Request The raw request sent to the provider.
Response The assembled response.

Note

Without the "See runtime-only and hidden site content" permission (site.hidden.view), the preview and the inspector show protected site content redacted. See Template reference.

Prompt fields

Field What it does
Owner Shown only when you create a prompt and may manage site content. Personal (only me) keeps it yours; Site (every user) makes it a site prompt everyone can use.
Name The prompt's name. Names are unique per owner.
Description Notes for yourself, in Markdown. Not sent to the model.
Purpose What the prompt is for (see the next table). Only Roleplay prompts can be bound.
Thinking How hard a reasoning model thinks for this prompt. Model's level uses the model's own Thinking setting. A level here is never higher than the model's own level; on a model left at Backend default it is sent as it is. Off suits mechanical work such as compaction.
Enabled A prompt that is off is never picked, whatever its bindings.
Visible to users Site prompts only. Everything lets users read the blocks; Only names (runtime-only) shows users only the names, never the content of its blocks, in the prompt inspector, the assistant, scripts or exports. The model still receives it, and admins always see everything.
Tags Tags for finding and filtering prompts (see Tags).

Purposes

Purpose Used for
Roleplay Chat replies. The only purpose that is bound.
Memory Writing a memory from a scene of the chat.
Consolidation Merging memories into a higher tier.
Compaction Shortening an existing memory entry.
Topical clip Writing a clip about one topic.
Side prompt Updating a side-prompt entry.
Clip review, Clip suggestions, Combine clips Memory Assistance jobs.
Caption Describing an image; chosen on a vision model.

The memory purposes are chosen by memory profiles and side prompts rather than bindings (see Memory), and caption prompts by the model that describes images (see Images and attachments). The Type filter above the prompts list narrows it to one purpose; All types shows every prompt.

Site and personal prompts

Prompts, blocks and outlets are either site items, which every user can use and which only users with the site.content.manage permission edit, or personal items, which only their owner sees, edits and uses. Creating personal ones needs the content.prompts.create permission, and the quota.prompts quota can cap how many prompts you own (see Users, groups and permissions).

A personal prompt can include site blocks and your own blocks, and can bind only to site or your own connections, models and characters, and to your own chats. You can also export a prompt, share it with other users, or publish it to the catalog from its editor (see Sharing and the catalog).