SillyTavern Prompt Itemization: Find What Uses Your Context

SillyTavern Prompt Itemization shows how the context for an AI reply is divided between instructions, character information, World Info, chat history, and extensions. Open it when a conversation seems to forget earlier events or a large character card leaves little room for dialogue. Then read the raw prompt to check the specific text involved.1

The breakdown answers “where did the space go?” Reading the prompt answers “was this sentence included?” Keep those questions separate before changing temperature or rewriting the character.

This guide follows the official documentation and release-branch source reviewed on 2026-11-03. It is a source-based diagnostic, not a report of a running SillyTavern test. UI details can differ by release and API.

Open the breakdown for the reply you are investigating

The short version: Generate a reply, expand its ... message menu, and choose the lined-square Prompt Itemization icon.

That is the route described in SillyTavern’s Common Settings documentation.1 Select the AI reply associated with the problem, rather than assuming another message represents the same request.

Before editing anything, record the API/model, preset, and tokenizer shown at the top of the panel. Save the current context and response limits too. These details matter when comparing token estimates across requests.3

Prompt Itemization is associated with stored prompt data for a message. The implementation searches for a matching message ID and returns without opening a breakdown if none is found.5 If an older message has no usable breakdown, inspect a newly generated reply in a disposable copy of the conversation. Do not treat missing diagnostic data as proof that the original request contained no character or lore information.

The chat’s dotted context-range line is another clue: the documentation says messages above it are not sent to the AI.1 It is useful for spotting history loss, but it does not tell you which non-history section consumed the space.

Read the totals before the individual fields

The short version: Use the estimated total and available prompt allowance first; do not add every visible row together.

The current templates show Total Tokens in Prompt and Max Context (Context Size - Response Length). The Text Completion template also includes Padding and Actual Max Context Allowed. Both templates explicitly warn that all numbers are estimates and that grey items may not have been included because of prompt-format settings.3

Read these as diagnostic estimates, not a provider’s billing statement. A field containing text in an editor does not establish that the active formatting sent it.

The layout also includes parent sections with indented details. Extensions, for example, has rows such as Summarize, Author’s Note, and Vector Storage (Chats). Adding a parent total and all its component rows would count the same material more than once. The implementation calculates totals differently for Chat Completion and Text Completion, so a homemade sum of every displayed number is not a reliable cross-check.35

Look at the large categories, then inspect their details:

These are leads, not reasons to delete the largest section. A long recent conversation can be more useful than a short generic instruction. The question is which text earns its space for the next reply.

Use Show Raw Prompt to check inclusion and duplication

The short version: Search for the needed fact in the prompt text, then check its surrounding material and origin.

The release template labels the inspection controls Show Raw Prompt, Copy Prompt, and Show Prompt Differences.3 Show Raw Prompt displays stored prompt text; Copy Prompt lets you save it privately for comparison. The differences control is available when the implementation finds a preceding stored raw prompt.5

Search for a distinctive sentence rather than a common character name. Suppose the important fact is The north gate closes at sunset. Finding north gate in your own latest question does not establish that a lore entry supplied the closing time. Finding the complete fact only in an old assistant reply shows a different route into context than a fresh World Info insertion.

Also search for repeated instructions. The same writing rule might occur in a character description and an Author’s Note. Decide whether the repetition is intentional before removing it. Token space alone cannot tell you which instruction matters more to the scene.

For Chat Completion, there is an important display limit: the current raw-view and copy code joins message content into text.5 That flattened view is useful for sentence searches, but it is not a complete representation of message roles. If the problem concerns role placement, inspect structured prompt data rather than inferring roles from paragraph order.

Keep copied prompts private. They can contain personal chat history, character text, and retrieved documents that were not visible in your latest message.

Separate missing content from content the model ignored

The short version: Classify the failure before choosing a setting to change.

Use the inspected request to choose the next investigation:

If World Info content is missing, follow the separate SillyTavern lorebook diagnostic. That guide covers assignment and activation. A World Info token total here cannot identify which particular entry succeeded.

If the needed information is present, changing token allocation may not address the problem. Try a short question about that fact in a disposable conversation. Only after checking the supplied content should you move to controlled roleplay LLM settings comparisons. Temperature cannot restore a sentence that was never sent.

For prompt placement changes, use the correct editor: SillyTavern documents Prompt Manager for Chat Completion and Advanced Formatting for Text Completion.2 Do not apply a Chat Completion walkthrough to a differently formatted backend.

A context ledger shows what a smaller prompt could buy

The short version: Reducing non-history content creates capacity for conversation, but verify which messages actually return.

The following is an illustrative capacity calculation, not measured output or a transcription of the panel. Assume an 8,192-token context setting and a 1,024-token response reservation, leaving 7,168 tokens for the prompt. For this simplified ledger, set aside tokenizer differences and padding.

Imagine inspecting the baseline raw prompt and finding a long background passage duplicated in two sources. In a copied setup, shortening that duplication by an illustrative 700 tokens leaves 700 more tokens available for history. It does not increase the model’s context window or promise longer replies.

That capacity is not a guarantee that a particular older message will fit. Check the next request’s history and search for the event you needed. Do not claim success merely because the total dropped.

Save a baseline prompt and make one source edit. Restore the same pre-reply conversation state before generating the comparison; otherwise, the baseline answer itself becomes extra history. Keep the model, preset, response limit, and other prompt sources fixed. Record which sentences disappeared and which history became available, not just whether the new answer sounded better.

If the shortened passage contained an important relationship fact, restore that fact in a concise form. The goal is useful context, not the smallest possible token count.

Use Prompt Inspector when the breakdown is insufficient

The short version: Prompt Inspector is a separate pre-send inspection tool, with limits that matter during diagnosis.

The official Prompt Inspector extension adds a wand-menu toggle. Once enabled, sending a chat prompt opens a popup containing prompt data. Its README describes Chat Completion data as a JSON-serialized array of objects before backend-specific post-processing; Text Completion uses a string.7

That makes it useful for examining structure that the flattened raw-text view does not preserve. It still does not prove what provider-side processing might add or change.

Inspect without experimenting on your main conversation. The README warns that Cancel discards changes but does not cancel the request. OK sends the modified prompt, and those modifications are temporary rather than saved configuration.7 An edited inspector request therefore does not prove that you fixed the underlying character field or preset. Make the lasting change at its source and inspect another request.

This procedure concerns SillyTavern. It makes no claim that Local Waifu imports its prompts or duplicates these controls. For a broader app decision rather than a prompt-budget fault, see the SillyTavern and Local Waifu comparison.

Sources

The short version: The UI route comes from official docs; field names and raw-view behavior come from the release source, not an invented walkthrough.

1. SillyTavern Common Settings: context range and Prompt Itemization.
2. SillyTavern Prompts: prompt components and inspection methods.
3. Chat Completion itemization template: labels and estimate warning.
4. Text Completion itemization template: labels and padding.
5. Itemized prompts implementation: message lookup, calculations, copying, and differences.
6. SillyTavern Prompt Manager: Chat Completion prompt order and roles.
7. Official Prompt Inspector README: workflow and limitations.

Questions people ask

Back to the blog