An LLM-powered extension for analyzing, scoring, and iteratively improving character cards.

Stop guessing if your character card is good. Let an AI tear it apart, tell you exactly what's wrong, rewrite it better, and then verify it didn't lose the character's soul in the process.

![Main Interface Screenshot Placeholder]

Character Tools runs your character cards through a three-stage pipeline:

1. Score — Get brutal, field-by-field feedback on what works and what's garbage
2. Rewrite — Generate an improved version based on that feedback
3. Analyze — Compare the rewrite against the original to catch "soul loss" and regressions

Then iterate until it's actually good, or accept and move on.

Not just "this card is a 7/10" — you get specific ratings and feedback for each field: description, personality, first message, scenario, example messages, system prompt, and more.

![Score Results Screenshot Placeholder]

The rewrite stage uses your score feedback to generate improvements. It knows what was weak and targets those areas while preserving what worked.

The analyze stage compares your rewrite against the original and answers the critical question: Does this still feel like the same character?

• What was preserved
• What was lost
• What was gained
• Verdict: ACCEPT, NEEDS_REFINEMENT, or REGRESSION

![Analysis Results Screenshot Placeholder]

If the analysis says "needs work," hit Refine. The extension:

• Saves your current iteration to history
• Generates a new rewrite addressing the identified issues
• Lets you analyze again
• Repeat until it's right

Full iteration history with one-click revert to any previous version.

![Iteration History Screenshot Placeholder]

Don't want to rewrite the whole card? Select only the fields you care about. Working on alternate greetings? Pick specific ones.

![Field Selection Screenshot Placeholder]

Enable JSON schemas for consistent, parseable results. Built-in schemas for Score and Analyze stages, or create your own.

Don't know JSON Schema? No problem. Click Generate and describe what you want in plain English:

"rating 1-10, list of issues, summary, and a boolean for whether it needs more work"

The AI builds the schema for you.

![Schema Generation Screenshot Placeholder]

Save your prompts and schemas as presets. Includes sensible defaults, but you can customize everything:

• Stage-specific prompts
• Custom JSON schemas
• System prompt additions
• Refinement instructions

• Chat Completion APIs — OpenRouter, OpenAI, Anthropic/Claude, Google AI Studio, Mistral, Groq, etc.

• Text Completion APIs — Might work, might not. We haven't tested it. If you try it and it works (or explodes), let us know.

The extension uses SillyTavern's generateRaw function, which should work with text completion, but structured output and the prompt format may behave differently.

1. Open SillyTavern
2. Go to Extensions panel (stacked cubes icon)
3. Click Install Extension
4. Paste: https://github.com/Inktomi93/SillyTavern-CharacterTools
5. Click Install
6. Refresh your browser

cd SillyTavern/data/<your-user>/extensions/third-party/
git clone https://github.com/Inktomi93/SillyTavern-CharacterTools

Restart SillyTavern after installation.

Find Character Tools in the Extensions panel, click Open Character Tools.

![Extension Panel Screenshot Placeholder]

Search for a character by name. The extension shows all populated fields with token counts.

![Character Search Screenshot Placeholder]

By default, all populated fields are selected. Uncheck any you want to skip. For alternate greetings, you can select specific ones.

The pipeline has three stages: Score → Rewrite → Analyze

• Check/uncheck stages to include them
• Click a stage button to view/edit its configuration
• Each stage has its own prompt and optional JSON schema

![Pipeline Nav Screenshot Placeholder]

• Run Stage — Run just the currently selected stage
• Run Selected — Run all checked stages in sequence

Results appear in the panel below. For each stage you can:

• Lock — Prevent accidental regeneration
• Copy — Copy to clipboard
• Regenerate — Try again with the same settings
• Continue — Move to the next stage

![Results Panel Screenshot Placeholder]

After Analyze, if the verdict is NEEDS_REFINEMENT:

1. Click Refine to generate an improved rewrite
2. Click Analyze This Rewrite to check the new version
3. Repeat until you get ACCEPT or decide it's good enough
4. Click Accept Rewrite to lock it as final

Click Export to download a markdown file with:

• All stage results
• Full iteration history
• Timestamps and metadata

Input: Your character card (selected fields only)

Output: Field-by-field analysis with:

• Numerical rating (1-10)
• Strengths
• Weaknesses
• Specific improvement suggestions
• Overall score and priority improvements

When to use: Always start here. Even if you think your card is good, the score gives you a baseline and identifies blind spots.

Input: Original character + Score feedback (if available)

Output: Complete rewritten character card

When to use: After scoring, or standalone if you just want a fresh take. The rewrite incorporates score feedback automatically when available.

Input: Original character + Current rewrite

Output: Comparison analysis with:

• What was preserved from the original
• What was lost (the "soul check")
• What was improved
• Verdict: ACCEPT / NEEDS_REFINEMENT / REGRESSION
• Specific issues to address (if refinement needed)

When to use: After every rewrite. This is your quality gate.

The real power is in iteration:

Score → Rewrite → Analyze
                    ↓
              NEEDS_REFINEMENT?
                    ↓
                 Refine → Analyze → ...
                    ↓
                 ACCEPT?
                    ↓
                  Done

Each iteration is saved. If iteration #3 is worse than #2, revert with one click.

Access settings via the ⚙️ icon in the popup header.

![Settings Modal Screenshot Placeholder]

• Use Current SillyTavern Settings (recommended) — Uses your active API connection
• Custom Generation — Override with specific source, model, temperature, etc.

The system prompt is sent with every generation. You can:

• Add your own instructions (appended to the base)
• Edit the base prompt (advanced, affects all stages)

Instructions for the refinement loop. Customize how the AI approaches fixing identified issues.

• View all prompt and schema presets
• Delete custom presets (builtins are locked)
• Export/import custom presets for backup or sharing

You have two options:

Click the Generate button under the schema textarea and describe what you want:

"scores for description, personality, and first message (1-10 each), a list of the top 3 problems, and an overall recommendation"

The AI creates a valid JSON schema from your description. Review it, tweak if needed, save as a preset.

If you know JSON Schema, write it directly. The extension validates as you type and warns about compatibility issues with different providers.

Required format:

{
  "name": "MySchema",
  "strict": true,
  "value": {
    "type": "object",
    "additionalProperties": false,
    "properties": {
      "score": { "type": "number" },
      "issues": { "type": "array", "items": { "type": "string" } }
    },
    "required": ["score", "issues"]
  }
}

Use Validate to check for errors, Auto-Fix to add missing additionalProperties: false, and Format to prettify.

1. Use a capable model — Claude, GPT-4, or similar. Smaller models may struggle with nuanced analysis.

2. Don't skip Analyze — It's tempting to just take the rewrite, but the soul check catches problems you won't notice until you're mid-roleplay.

3. Trust REGRESSION verdicts — If the AI says it got worse, it probably did. Revert and try different refinement instructions.

4. Iterate 2-3 times max — If it's not converging after 3 iterations, the original might need manual work first.

The built-in prompts are solid defaults, but you can:

• Create stage-specific presets for different character types
• Add constraints ("keep it under 500 words", "maintain first-person perspective")
• Use placeholders like {{original_character}} and {{score_results}} for precise control

JSON schemas force consistent output format. Good for:

• Programmatic processing of results
• Ensuring the AI doesn't skip sections
• Getting machine-readable scores

Not all models support structured output well. If results look wrong, try disabling it.

• Refresh your browser
• Check it's enabled in the Extensions panel
• Look for errors in browser console (F12)

• Verify your API is connected (green dot in popup header)
• Check SillyTavern console for error details
• Try a different model if one is consistently failing

• Disable Structured Output if your model doesn't support it well
• Try a more capable model
• Check that your prompts aren't conflicting with the base instructions

• Make sure you've searched and clicked on a character
• The character must have at least one populated field

• This extension was built for Chat Completion. Text Completion might work but is untested.
• Try switching to a Chat Completion source if available.

• SillyTavern 1.12.0+
• A connected LLM API (Chat Completion recommended — OpenRouter, OpenAI, Claude, etc.)
• A model capable of following complex instructions (Claude Sonnet/Opus, GPT-4, etc. recommended)

AGPL-3.0

GitHub Issues