⚒ PresetForge

A generative, visual builder for SillyTavern preset kits. It treats a preset as a program — modules, variables, regexes, automations — and gives each one an editor, a linter, and an LLM that never touches the JSON.

Plan → · SillyTavern → · Run it ↓

---

Note

The LLM never writes preset JSON. It writes module content and structured plans; the app owns every identifier, marker, and prompt_order entry. That one decision deletes the entire "Invalid file" failure class that hand-authored and chat-generated presets suffer from — and it is why every export here passes an independent validator.

🧭 The idea

SillyTavern's popular presets stopped being settings files years ago. NemoEngine 11.5 is 451 prompt modules and 91 regex scripts; Chimera is 90 modules with 30 in-chat injections; whole browser extensions exist just to navigate presets this size. A preset like that is a small program — with state, load order, mutually-exclusive branches, output post-processing, and automations riding alongside. It is edited today through a drag-and-drop list or raw JSON.

PresetForge is the missing IDE: import anything, see what the model will actually receive, change it with confidence, generate the boring parts, and ship the result as a distribution-grade kit.

flowchart LR
    subgraph AUTHOR["🧠 AUTHOR — humans and models"]
        direction TB
        W["<b>Wizard</b><br/>plan → approve → generate<br/><i>expand mode reads your preset</i>"]
        QR["<b>Quick Replies</b><br/>STScript from a description"]
        AD["<b>🎯 Advisor</b><br/>reads a character card,<br/>recommends toggles"]
    end

    subgraph FORGE["⚒ FORGE — the working model"]
        direction TB
        ED["<b>editor</b><br/>modules · variables · regexes"]
        LN["<b>lint</b><br/>format traps · 🔗 conflicts"]
        PV["<b>context preview</b><br/>where prompts actually land"]
        ED --> LN --> PV
    end

    subgraph SHIP["📦 SHIP"]
        direction TB
        K["<b>kit export</b><br/>preset + README + GUIDE<br/>+ regex/ + quick-replies/"]
        ST["<b>ST link</b><br/>save straight into<br/>SillyTavern's folder"]
    end

    W --> ED
    AD --> ED
    QR --> K
    PV --> K
    PV --> ST

    style AUTHOR fill:#221a33,stroke:#8b5cf6,color:#eee6ff
    style FORGE fill:#16242d,stroke:#3d9ae8,color:#e6eef0
    style SHIP fill:#1a2d1f,stroke:#3dc26f,color:#e8f5ec

Loading

---

🗂 The workspace

A preset library, not a single file. Slots switch instantly from the toolbar; imports — from disk or straight from a URL (🌐, GitHub raw links work) — always open a new slot, so checking out someone else's preset can never overwrite an hour of your work. Duplicate a preset before surgery, delete the experiments after.

🕒 Snapshots back every slot: a ring of five per preset, taken automatically before wizard runs and restores, or manually whenever. The diff view shows exactly what changed between then and now — modules added/removed/edited, toggles flipped, sampler params moved, order rearranged — and restore keeps an escape-hatch snapshot of the state you're leaving.

At NemoEngine scale the outline stays navigable: divider modules become collapsible sections (both the ━━━ empty-divider convention and Nemo's === Banner === style are detected), collapse-all turns 451 rows into a 16-line table of contents, and Ctrl+K fuzzy-jumps to any module by name. Rows are memoized — a toggle re-renders one row, not four hundred.

📦 Module library: hit ☆ on any module to bank it as a reusable block, then insert it into any other preset with a fresh identifier. Your best genre toggles, CoT blocks, and stances stop being copy-paste jobs.

🤖 Authoring with models

Every LLM feature speaks to LM Studio, OpenRouter, OpenAI, or Anthropic — prebuilt service entries where the API key is the only thing you type. The model dropdown fetches live from the server's /models endpoint, which doubles as the connection test: models appearing is the proof it's wired.

• ✨ Wizard — describe the preset; the model returns a module plan (names, categories, placements, defaults) which you prune before any content is generated. Generation runs three requests in parallel and applies atomically — a failure at module 9 of 12 leaves your preset untouched. Expand mode points the wizard at an existing preset instead: it reads the current module names, plans only complements, and leaves your Main Prompt and samplers alone.
• 🎯 Advisor — load a character card (V2/V3 JSON, or a card PNG — the embedded chara/ccv3 chunk is parsed in-browser) and the advisor reads it against your module list, recommending per-module enable/disable with reasons, applied selectively in one click. NemoGuides' idea, moved to authoring time.
• Refine — every module editor carries an instruction box: "make it half as long and more forceful" returns a proposal you accept or discard. Proposals are keyed to the module, so switching selection mid-flight can never write the wrong module.
• ⚡ QR generation — see below; the STScript primer is LALib-aware.

🔬 The linter and the preview

The format knowledge comes from SillyTavern's source, not its docs or community lore — and that distinction has teeth. Things the lint catches that circulate wrongly in guides:

• presets wrapped in {"version","type","data"} (imports as a dead preset)
• flat prompt_order lists (silently ignored by ST — the real shape nests under character_id, and Chat Completion reads 100001)
• injection_depth set on Relative-position prompts (depth only applies In-Chat — entire "depth strategies" in popular guides do nothing)
• {{getvar::name::}} — the trailing :: breaks the lookup
• deprecated keys (claude_use_sysprompt → use_sysprompt), proxy keys that trigger import warnings, unreachable modules, duplicate order rows, enabled In-Chat prompts with Chat History disabled (they'd never be sent)

The context preview renders the assembled stack against a sample scene or your loaded card: relative prompts in order, markers expanded, In-Chat prompts spliced into the chat at their true depth — with same-depth ordering matched to populationInjectionPrompts in ST's source. Click any block to jump to its module.

Warning

The preview is structural, not byte-identical. It promises correct placement — which is the thing everyone gets wrong. Fun fact from building it: SillyTavern's own documentation describes same-depth injection role order incorrectly; the preview follows the source instead. Trust code over docs.

🧵 Variables

{{setvar}}/{{getvar}} networks get IDE treatment: the Vars tab maps every definition and read across the preset, renames a variable through all 90 modules in one collision-checked operation, jumps to any usage with the macro literally selected in the editor, and offers a one-click redirect for dangling reads pointing at variables nothing defines.

🧪 Regex kit

The big engines ship regex scripts because presets shape the input but regexes shape the output — em-dash removal, format unification, dialogue styling. PresetForge authors them natively: placement flags (AI output / user input / reasoning / …), display-vs-prompt targeting, per-script validation, and a live test box that runs your find/replace as you type. Scripts store inside the preset at extensions.regex_scripts — the modern ST mechanism, gated behind ST's own per-preset allow prompt — or export as standalone files for older setups.

⚡ Quick Replies and STScript

Preset kits ship automations too. The QR editor builds version: 2 Quick Reply sets (schema straight from ST's QuickReplySet.toJSON), lints STScript closure balance inline, and generates scripts from a description — "a d20 skill-check button, a rewind button, a scene summarizer" — against a primer of core STScript. Toggle LALib mode (on by default) and the generator may use LenAnderson's 98-command library: /= expressions, /foreach//switch control flow, list and dict ops, swipe and message surgery, even /qr-add.

🧩 Text Completion wing

For the local-model crowd: an instruct / context / system-prompt template editor with schema-driven forms verified against ST's shipped templates (ChatML, Llama-3, …). Import any template JSON — the kind is auto-detected — edit every sequence and flag, export it back.

📦 Shipping

• 🧰 Kit export writes a distribution folder in one click: the preset, a generated README.md (install steps including the regex allow-gate, defaults, exclusion groups, sampler table), a MODULE_GUIDE.md documenting every module grouped by section, standalone regex/ scripts, and quick-replies/ sets. Run against NemoEngine it produces 94 files and documents 422 modules.
• 🔗 ST link goes one better than export: pick your SillyTavern/data/<user>/OpenAI Settings folder once (the handle persists), then Save to ST writes directly into SillyTavern and Open from ST imports from the live folder. The export→import dance is gone.
• 📖 README generator builds the in-preset documentation module — the mega-preset convention — from the actual module tree, zero LLM required.
• 🔗 exclusion groups: name variants 🔗 POV: First Person / 🔗 POV: Third Person and enabling one auto-disables its siblings; imports that arrive with conflicting groups enabled get flagged by lint instead.

📏 The verification story

This tool edits 451-module programs, so it doesn't get to guess:

	claim	receipt
1	Round-trips are lossless	Test suite asserts semantic identity — params, prompts, order, per-character entries — on real NemoEngine and Chimera files; exits nonzero on drift
2	The format model is true	Every schema fact verified against ST 1.18 release source; the sibling creating-sillytavern-presets skill shares the same rule set
3	The code was hunted	Eleven adversarial review agents before v0.2: 15 findings reported, 15 fixed — including the preview's injection order being wrong on both axes
4	It holds at scale	451-module import in ~5ms, ~41ms per edit frame at 453 modules, sections collapse the wall to a table of contents

🧰 Run

npm install
npm run dev        # http://localhost:5299

Everything is local-first: presets, the module library, TC templates, and QR sets persist in the browser; API keys never leave it. No backend, no build step beyond Vite.

npx esbuild tests/roundtrip.ts --bundle --format=esm --platform=node \
  --outfile=/tmp/rt.mjs && node /tmp/rt.mjs [path-to-preset.json]

🙏 Lineage

This tool stands on other people's work, and says so:

• cha1latte's sillytavern-preset-creator — the seed. Its three-part guide to AI-authored presets shaped this app's module patterns, tier thinking, and defaults strategy — and chasing its format claims down against SillyTavern's actual source became the founding discipline here: verify everything, trust code over lore. This project exists because that guide asked the right question first.
• NemoVonNirgend — NemoEngine is the stress-test fixture and the proof that presets became programs; NemoPresetExt's conventions inform the section system, and NemoGuides' Prompt Advisor inspired 🎯.
• LenAnderson — LALib's 98 commands power the QR generator's richest mode. H A I L Lenny.
• Nativu5's STPresetEditor — prior art for visual preset editing; its variable find-usages set the bar the Vars tab aims at.
• The SillyTavern team — the platform all of this orbits.

📁 Layout

preset-forge/
├── src/lib/
│   ├── preset.ts        normalize/export -- the only code allowed to touch JSON shape
│   ├── lint.ts          the trap detector; rule parity with the Python validator
│   ├── assemble.ts      context preview -- placement per ST source, not docs
│   ├── macros.ts        setvar/getvar graph, rename machinery
│   ├── groups.ts        🔗 exclusion groups + section-header detection
│   ├── regex.ts         ST regex scripts (extensions.regex_scripts)
│   ├── quickReplies.ts  QR sets + STScript closure linting
│   ├── tcTemplates.ts   instruct/context/sysprompt schemas
│   ├── snapshots.ts     per-preset ring buffer + structural diff
│   ├── kit.ts           preset + README + MODULE_GUIDE + kit folders
│   ├── stLink.ts        FS-Access bridge into SillyTavern's own folder
│   ├── cards.ts         V2/V3 card JSON + PNG chara-chunk parsing
│   └── gen.ts           every LLM prompt in the app lives here
├── src/components/      one file per pane/modal, no framework beyond React
├── tests/               round-trip + feature + kit suites against real presets
└── PLAN.md              phases, verify gates, the parking lot

---

Verified against SillyTavern 1.18's source, stress-tested on the biggest preset in the wild, and reviewed by eleven adversarial agents before v0.2 — because a tool that edits 451-module programs doesn't get to guess.

⚒ Plan · 🧪 Tests · 🐙 SillyTavern