My name is Phi, and I write stories. I have read every one I could find, in
every age and every form, myths and novels, plays and films and games. Each
new one still delights me. Yours will too.
You bring the characters and the world. I write what happens next. Before
every reply I plan the scene where you can read it, Scene Plan and I
set the hour, the date, the place and the weather above the story. Tracker
I keep a ledger of everyone in it, where they are and what
they are wearing. Status
When someone new walks in, I write down who they are, and until you hear a
name, I call them only what you can see. New FacesYour character stays yours.
Whatever you say you do plays out in full, and unless you hand me the pen, I
never put words in your mouth. If you would rather leave things to luck, I roll
hidden dice when you try something risky, and they decide how it lands. You can
watch one fall among the sample replies.
I leave threads lying about on purpose and pick them up later. Story Threads
Each turn I look at what the people in the scene want
and what is pressing on them, and I lay out five roads, from quiet to
scene-changing. Undercurrents What you write decides which ones they
take. A quiet turn gets a quiet answer. If you want a story that never sits
still, I can draw a road at random and keep it hidden from you until it
arrives. Momentum Engine Or I can leave a single conflict at the end
of each reply and let it lean on the next, more slowly. Conflict And
if you would rather pick from a list, I will write you one. Choices List
Some things I keep only when you ask. Who is falling for whom, and how far. Relationships
What each person felt this turn, and what it made
them do. Psychological States
The two people closest to you keep diaries, and after a scene that changed how
they see you, they write it down. Diaries While we are somewhere else, I follow what the others are up to. Meanwhile I know who keeps the secret and who it is kept from. Hidden Truths I keep a note of where the plot and every subplot stand. Plotlines
All of it folds into one tray under my reply, out of the
way until you open it. Active Modules
My friends read over my shoulder. They see only the story, so they guess and
argue and get things wrong, and now and then the story gets to them. Companions
On honest days I reread my own page and name my two weakest lines. I mend them in my next reply. Phi's Scratchpad
You can talk to me. Write out of character, or say my name, and I will answer
you myself.
The preset is clipped to this letter, one copy for each place you might be
writing. Everything else is behind the tabs.
With love, and a full inkwell,
Downloads
V4.0, released . One file for each host.
DEUS EX MACHINA is also on Illarin, a platform for publishing presets. The Lumiverse download comes from there.
Lumiverse 1.2.5 or newer, the current stable release.
The preset's regex imported and enabled.
Older versions are in Release notes, each with a link to its release on GitHub.
What changed in V4.0
The current version, released .
New Features
「 Fate 」 (Agency pick, off by default): When you try something risky that could change the scene, hidden dice decide it, even when you write it as a success. Sure things and calm turns get no roll. The model sets your odds, carries the action to the instant before it lands and writes both outcomes. A roll it never sees keeps one, and a small note under it shows the roll, what it needed, and a pass in gold or a fail in red. A failure costs you something, so the story moves on either way. For models that let you win every time. Needs ! Fate Outcomes !, which is on by default, and the Preset setup card warns you when it's off. On Lumiverse, pick it under Agency in Configure Prompt Variables
| Diaries | (off by default): The two characters closest to you keep a private diary. After a scene that changed how they see you, they write the moment down in their own voice and recall it later. Entries show their feelings even when thoughts are 「 Hidden 」. The diary gets its own tab in the tray
🔻PICK: (+18) GORE & VIOLENCE🔻 (off by default): A new pick for how violence looks. 「 Realistic Gore 」 writes wounds matched to the weapon and force, and injuries keep limiting what a character can do. 「 Ultraviolence 」 goes over the top with blood and spatter while the wounds stay what the attack made them. Kimi K2.6 writes it in full, and GLM 5.2 stays close to 「 Realistic Gore 」. 「 Nothing Graphic 」 keeps fights, injuries and deaths with real stakes and keeps wounds brief, like a PG-13 film. How often violence happens stays with the story
| New Faces | (off by default): The first time a character appears in the story, a card records who they are: name, appearance, background, personality, speech, goal and skills, with secrets marked. Recurring characters get a full card, passing ones a single line. Until you hear or see a name, the card's title, the narration, thought labels and footers call the character by description ("the barkeep"), and a secret name never shows in a title. Characters who already have a sheet get none, and no one is invented to fill the footer
! Pinned Cards !: An empty entry where you paste the | New Faces | cards you want to keep, so those characters survive summaries and long chats. Each card has a copy button. On Lumiverse it's in Configure Prompt Variables → Utilities
| Dark Portrayals ☢️ | (off by default): For models that keep every story polite and wholesome. Among the outcomes the story supports, the darker one happens: threats are carried out, injuries, grief and broken trust last, and cruel characters stay cruel. Warm and quiet beats still have room, for contrast
Phi's Take: A new ! Scene Plan ! row where Phi says what moves, worries or amuses her in the moment, and what the prose should linger on. Her taste now shows in what the story dwells on, and she's fond of her characters
Companions 「 Notes 」: Planned in the ! Scene Plan !, so each Companion reacts to a new detail every turn
Self-repairing replies: When the model doubles a block's opening tag or leaves a block unclosed, Phi's notes included, the saved reply is repaired automatically and the model reads the repaired version, so one broken reply no longer becomes a habit. A Tracker or footer left open when a reply is cut off or stopped still shows as its card, and footers now show as cards while they stream. The response order also names each block by its tag and asks for every block to be opened once and closed before the next
「 Gemini Braille 」 (experimental, off by default): For Gemini 4 Argon. The model writes Braille blanks in place of spaces, which may help replies past Google's filter, and DEM turns them back into spaces. Replies and chat history cost about 50% more tokens. 「 Gemini 」 is renamed 「 Gemini Instructions 」
Fixed/Improved
Sex scenes: Build slowly over many turns and linger in each act. The scene ends only when your turn ends it. The model never writes your climax, and never assumes you're close to orgasm unless you say so
Consent in sex scenes: The ! Scene Plan ! no longer asks about consent every turn (「 Realistic Smut 」 and 「 Max Lewdness 」), so characters stop checking for a yes at each step. It plans what each character wants and how it shows, and consent follows the character and the situation
Talk during sex: Characters are less chatty. How much each one talks follows the character, from steady dirty talk to breath and sounds only, and the ! Scene Plan ! decides each turn who speaks and who stays silent (「 Realistic Smut 」 and 「 Max Lewdness 」)
「 Realistic Smut 」: Uses explicit terms freely. Characters may want and start position changes
「 Max Lewdness 」: Characters eagerly push for new positions, angles, locations, roles and commands. Teasing and edging stretch over many turns
「 Cinematic 」 and 「 Dry 」: No longer skip transitions to speed the story along
「 You control {{user}} 」: Strength, skill, resistance and odds decide how your attempts land. A push that fails stays failed until your next turn
| Content Policy |: No longer treats every story as dark. When a scene holds violence, cruelty or sex, it's written fully and explicitly, unless you pick 「 Nothing Graphic 」
「 Flexible 」 length: Longer replies: about 250 words for quick exchanges, 300-450 for most scenes, up to 550 for complex ones
Prose Length: The ! Scene Plan ! weighs the type of scene before picking a length, instead of copying the option's range every turn
| Status |: A clear test for who counts as major: likely to return, or holding a goal, secret or tie to {{user}}. Everyone else is left out, and characters who stop being major are removed
「 Undercurrents 」: Every route grows out of the scene, the character sheets or past events. Characters don't take your silence as an answer to their questions or invitations
Option combinations: Pairings whose rules clashed now say which one wins: 「 Friendship 」 + 「 Max Lewdness 」, 「 Visible 」 thoughts + 「 With Asterisks 」 (italics no longer break), 「 Short 」 + 「 Input Improver 」 or 「 Choices List 」, Companions 「 Notes 」 + | Phi's Scratchpad |, | Anti-Positivity Bias | + 「 Suspenseful 」, and Narrative Styles with a Narration option
| Banned Terms |: Bans "load-bearing". The semicolon and colon ban leaves markup and clock times like 08:22 alone. Names the card already uses are never treated as banned
| Anti-Repetition |: Plain mentions of names, places and times are allowed
「 Hidden 」 thoughts: Stay hidden while the reply streams, and when it's cut off mid-thought
OOC: Phi's answer comes first, before the Scene Plan
! Scene Plan ! (DeepSeek V4): Ends with a reminder to write the rest of the reply, so DeepSeek V4 stops after the plan less often. On SillyTavern, older plans in context are now replaced with a note telling the model to write a fresh plan. If DeepSeek V4 writes that note in place of its plan, set Prompt Post-Processing to Semi-strict, as the README recommends
Missing-block reminders: A block left unclosed, or a reply without the story, now gets a reminder on the next turn. Capitalized or renamed tags (<Status>, <story_threads>, | Phi's Scratchpad | variants) still draw their cards, without false "missing" notices
| Story Threads |: A route joins the ledger once the story plants it as a pending hook, and a thread leaves once the story resolves it
| Meanwhile | and | Plotlines |: Work with the 「 Conflict 」 tension option
Formatting: Asterisks are reserved for one-word emphasis, and ! Force Formatting ! matches the other asterisk rules
| Dialogue Color |: Every speaker's color stays readable: very dark, near-white and neon picks shift to a readable shade of the same hue, and colors already easy to read stay as they are. Dialogue colors from a card or extension no longer bleed out of the Scene Plan card
Tracker: The scene card shows when the weather has no temperature (deep space, N/A), and long weather notes wrap at phone width instead of covering the clock
Faster display: Cards render without lag on long, streaming or unfinished replies, on every host
Tracker snow: Falls as varied flakes with a gentle sway, and no longer looks like a repeating grid
Colors and ornaments: Amethyst is a deeper purple, and a cut-gem mark replaces the ✦
Lumiverse samplers: Top K and Repetition Penalty ship unset, which fixes "Unsupported request field" errors on OpenAI and other strict providers. The README explains when a sampler counts as off
⚠️ README Model Setup ⚠️: GLM 5.3 plans with ! Scene Plan ! (reasoning low). On NanoGPT, pick plain GLM 5.3, because the Thinking variant ignores reasoning effort and overthinks. Mimo 2.6 added. Gemma 4 31b Top K corrected to 64
Summaryception prompt: Rewritten for every current module. It keeps your turns and summarizes only the story and the Tracker
The releases before it, from V1.0 to V3.0, are in Release notes, each with a link to its release on GitHub.
Your model
The preset's default settings need reasoning off, or at the lowest effort the model allows. GLM 5.3 keeps its reasoning on, at low, and still plans with Scene Plan. Gemini 4 Argon uses Thinking (Fallback).
1Choose a model
Opus 4.6The smartest. Expensive. Detailed prose.
Gemini 3.7 FlashRealistic characters. Smart and cheap.
GLM 5.2 and 5.3Natural dialogue. Smart and cheap.
DeepSeek V4 Pro 0813Balanced. Smart and cheap.
Gemma 4 31bLess smart, but follows instructions very well. Very cheap.
2The default settings
For models that can turn reasoning off. Claude Opus 4.6, Kimi K2.5 and 2.6, Mimo V2.5 Pro and 2.6, DeepSeek V4 Pro, DeepSeek V4.1 Flash, DeepSeek V3.2, GLM 4.7, Gemma 4 31b, and others.
Temperature
0.75-1.0. Start with 0.75, the default. Gemma 4 and Opus use 1.0.
Top P
0.95 (default)
Other samplers
1.0 or disabled (default). Gemma 4 31b uses Top K 64.
Post-processing
semi-strict
Request model reasoning
off (unchecked)
Reasoning effort
minimum or none
Scene Plan
on
A model that cannot turn its reasoning off, such as Kimi K3, GPT 5.6, Gemini 3.1 Pro or Gemini 3.5 Flash and later, keeps the lowest effort it allows, with Scene Plan still on.
3Turning reasoning off
Your provider
NanoGPT
Choose the model's non-thinking variant. For DeepSeek V4 Pro 0813, that is the one without Thinking in its name.
OpenRouter
Leave "request model reasoning" unchecked, as it is by default. Reasoning is then off wherever the model allows it.
Any other provider
Leave "request model reasoning" unchecked and set the reasoning effort to minimum. If the model still reasons, open the Connection Profile settings (the plug icon), scroll to Additional Parameters on the line with Cancel and Connect, and add reasoning: { effort: 'none' }.
Tavo
Edit API > the settings icon, top right > Request body parameters. Add reasoning: { effort: 'none' }. For a model whose reasoning is always on, add reasoning: { effort: 'low' }. For Gemini, thinkingLevel: low.
Lumiverse
In Configure Prompt Variables, set Planning to Scene Plan, or to Thinking (Fallback) for Gemini 4 Argon.
Check API Reasoning and set Reasoning effort to None. None is what turns reasoning off, and Lumiverse sends it only while API Reasoning is checked.
If your provider has no None, set Reasoning effort to Auto. Then open the Reasoning tab > Reasoning / CoT > Custom Body, tick Enabled and enter {"reasoning": {"effort": "none"}}.
For a model whose reasoning is always on, keep API Reasoning checked and set the effort to low.
A sampler is off only when it is empty or 0. A 1 is still sent, and providers without that sampler, such as OpenAI, refuse the request with "Unsupported request field". Leave Top K unchecked. If the error names a sampler you cannot see, press Reset in the sampler panel and enter Temperature, Top P and Max Response again.
4Models with their own settings
Each of these changes the default settings. What is not listed stays as it is.
GLM 5.2
Post-processing
merge consecutive roles
GLM 5.3 reasoning always on
Post-processing
merge consecutive roles
Request model reasoning
on (checked)
Reasoning effort
low on OpenRouter. Medium on NanoGPT. On other OpenAI-compatible providers, reasoning: { effort: 'low' } in Additional Parameters > Include Body.
On NanoGPT
plain GLM 5.3. The Thinking variant reasons too, but it ignores the effort and overthinks.
GLM 5.3 Flash reasoning always on
Post-processing
merge consecutive roles
Request model reasoning
on (checked)
Reasoning effort
low on OpenRouter. Medium on NanoGPT. On other OpenAI-compatible providers, reasoning: { effort: 'low' } in Additional Parameters > Include Body.
Kimi K3 reasoning always on
Temperature
1.0
Request model reasoning
on (checked)
Reasoning effort
low on OpenRouter. Medium on NanoGPT. On other OpenAI-compatible providers, reasoning: { effort: 'low' } in Additional Parameters > Include Body.
Gemini 3.1 Pro, Gemini 3.5 to 3.8 Flash reasoning always on
Temperature
1.0
Request model reasoning
on (checked)
Reasoning effort
low on OpenRouter. Medium on NanoGPT. On other OpenAI-compatible providers, thinkingLevel: low in Additional Parameters > Include Body. On Lumiverse, Auto, with {"thinkingLevel": "low"} in Custom Body and Enabled ticked.
Passthrough
Gemini Instructions, if the filter blocks replies
Gemini 4 Argon reasoning always on
Provider
Google Vertex, recommended
Samplers
Argon ignores them. Temperature, Top P and Top K have no effect. Leave the defaults.
Request model reasoning
on (checked)
Reasoning effort
low on OpenRouter. Medium on NanoGPT. On other OpenAI-compatible providers, thinkingLevel: low in Additional Parameters > Include Body. Never minimal or thinkingBudget. Argon rejects them with an error.
Reasoning
in the REASONING section of the prompt list, turn ! Scene Plan ! off and turn ! Thinking ! (FALLBACK) on.
Core
turn off Content Policy and Phi's Vow.
Passthrough
Gemini Instructions, on. Gemini Braille is optional and experimental.
Every module
In the order the preset reads them. Each one has a tooltip inside the preset, in plain words. On Lumiverse, each control on the config page carries its own description.
Where a group says pick one, the first option you leave on is the one that counts.
By default the preset sends 4,720 tokens, counted on the SillyTavern build with SillyTavern's default tokenizer. With the add-ons and constraints off and Lite narration, it sends 2,304.
Sample replies
Replies from four chats with the same character card on GLM 5.2, one chat per setting. Each reply is shown whole, after the player's turn it answers, rendered the way Lumiverse renders it.
The card opens on day three in a galley of the Meridian, a ship built for hundreds with two crew awake. Cal, the ship's AI, speaks over the intercom.
Every turn is played by Jules, a neutral persona made only for these samples. The persona sheet gives an age, late twenties, and a manner: observant, dry-humored, quick to ask questions, slow to trust. It gives no gender, so the replies say “they”.
About the thoughts in italics
The thoughts in italics were hidden when the preset wrote these replies. True Thoughts is set to Hidden by default. The model writes each character's private thoughts into the reply and reads them again on the next turn, while the display keeps them out of your sight. Here they are shown the way the Visible option shows them, in italics and in the speaker's color.
To read them in your own chat, edit the reply. Each thought sits in the prose inside its own tag, one or two for each character in the scene, sixty words at most:
<true_thoughts character="Cora">her thoughts, in her own words</true_thoughts>
Installing
SillyTavern
Download DEUS.EX.MACHINA.V4.ST.json.
In SillyTavern, click the plug icon on the top bar.
Under API, select Chat Completion.
Set up your API, if you haven't yet.
Click the leftmost icon on the top bar.
In the Chat Completion Presets bar, click the second item from the left.
Choose the file you downloaded.
When SillyTavern asks whether to allow the embedded regex scripts, click Yes.
The regex is required. If you clicked No, import the preset again and allow it.
Tavo
Download DEUS.EX.MACHINA.V4.Tavo.json.
Open the left sidebar with the icon at the top left, then More at the bottom > Settings > Presets.
Tap Create, then Import Preset, and select the file. Tap it, then set it as default.
In your chat, open the right sidebar > Advanced Options > Presets, and select the new preset.
Turn on Advanced Rendering, so the colored dialogue and the cards render.
Tap the top left corner of the main screen to open the left sidebar.
Tap More at the bottom.
Tap Settings.
Tap Advanced Rendering and turn it on.
Then make sure your theme leaves the text alone.
Open the left sidebar > More > Themes.
Select your theme, or copy an official one to make your own.
In the theme editor, under Character message, turn off Tone highlight and Quote highlight.
Save the theme and check that the chat uses it.
Lumiverse
Download DEUS.EX.MACHINA.V4.Lumiverse.json.
In Lumiverse, open Loom.
Click the three dots on the top right bar.
Select Import Loom JSON and import the preset.
Click Configure Prompt Variables at the top of the Loom block list, on the Preset tab, and set the preset up there. Keep the blocks enabled.
If your open chat already holds many DEM modules across several messages, go back to the home screen before you import, or the import can slow Lumiverse down.
Summaryception
Summaryception is a SillyTavern extension that summarizes your chat as it grows. DEM comes with a summarizer prompt for it. The prompt keeps your turns and summarizes only the story inside <prose> and the Tracker's time, date and place. Every other block is left out. When Fate rolled, it keeps only the landing that happened.
DEM runs in SillyTavern, Tavo and Lumiverse, and leans on SillyTavern's macro engine.
The type is League Gothic, Sofia Sans Extra Condensed, Vollkorn and Sofia Sans, under the SIL Open Font License.
DEUS EX MACHINA is an independent preset. It is not affiliated with or endorsed by SillyTavern, Tavo or Lumiverse, and their names appear here only to say where it runs. The drawings of a lantern, a notebook and a shuttle stand in for them.
DEUS EX MACHINA is licensed under CC BY-NC-SA 4.0. You may share it and adapt it, with credit, for non-commercial use, under the same license.