super lazy
Thinking full CP
Use Universe, GLM / DS, or GPT 5.5 when you want one big character-RP custom prompt with the thinking checklist and full tracker already inside.
Start here.
Pick by bot shape before model name. The full CPs stay here as super-lazy thinking builds. PREBUILT files are complete no-thinking builds. MODULES are for custom assembly.
route before copy
lazy no-thinking
PREBUILT
Use one complete no-thinking build when the model has no reliable hidden reasoning, or when long checklists cause identity loops.
custom path
MODULES
Use modules when you want to assemble a smaller or more specific prompt. Start with Core Skeleton, then choose the bot shell.
pov note
Gender lock
Defaults are WLW and FemPOV-friendly. MLM or male POV builds may need pronoun, relationship term, and role-logic review before publishing.
Copy-ready builds.
First row: updated super-lazy thinking CPs with villain integrity logic and known token counts. Second row: standalone no-thinking PREBUILT prompts, token counts left as placeholders for now.
complete copy surfaces
super-lazy thinking build
Universe
Default full CP for a regular character bot with stable {{char}}, strong POV lock, body mechanics, heat calibration, villain integrity, thinking checklist, and full ID/EGO/SUPEREGO tracker.
view generation settings
temperature0.81 / 2.0
top_p1.0 / 1.0
freq penalty0 / 2.0
pres penalty0 / 2.0
max_tokens0, model default
context28k tokens
These are starting dials, not part of the visible build decision.
view full prompt
super-lazy thinking build
GLM / DS
Full character-RP CP with stronger hidden-info guard and bracket emphasis. Use when the model reads concealed plans, private thoughts, or unrevealed backstory as if the scene made them observable.
view generation settings
temperature0.81 / 2.0
top_p1.0 / 1.0
freq penalty0 / 2.0
pres penalty0 / 2.0
max_tokens0, model default
context28k tokens
Chosen for godmod control first. Tune sampling after behavior tests.
why chinese brackets
Q: Why the Chinese brackets【】?
You'll notice this preset uses 【brackets like these】 around key rules. This isn't decorative. DeepSeek and GLM were trained on large amounts of Chinese-language data, and these brackets, called lenticular brackets, show up constantly in Chinese writing as emphasis markers, similar to how English uses bold or ALL CAPS. Because of this, both models treat text inside 【】 as higher-priority instructions compared to the same text without them.
Think of it like highlighting the most important lines in a textbook. The model doesn't skip the un-bracketed parts, but the bracketed ones get extra attention.
Q: Should I keep them if I'm editing the preset?
Yes, generally keep them on lines that define hard behavioral rules, things the model should never break. If you're mixing modules or writing your own rules, wrapping your most critical lines in 【】 can help them stick, especially on DeepSeek V4 and GLM 5.1.
That said, they're not magic. Don't bracket everything or the emphasis loses meaning. A good rule of thumb: if a line is something you'd bold in a document because the model keeps ignoring it, that's a good candidate for 【】.
Q: Do I need them for other models?
Not really. Claude and GPT don't gain much from them since their training data is mostly English. They won't break anything if left in, but they also won't help. If you're porting this preset to Claude or GPT, you can safely remove them and use XML tags alone.
One last thing: No preset works perfectly out of the box for every bot. Read through the rules, understand what each section does, and adjust to fit your need. A preset is a starting point, not a finished product.
view full prompt
super-lazy thinking build
GPT 5.5
Full character-RP CP with overexecution control and villain integrity logic. Use when strong instruction-following activates too many mechanisms in one turn.
view generation settings
temperature0.81 / 2.0
top_p1.0 / 1.0
freq penalty0 / 2.0
pres penalty0 / 2.0
max_tokens0, model default
context28k tokens
The build card keeps overexecution risk visible; sampling dials stay secondary.
view full prompt
no-thinking lazy build
Character no-thinking
Use for one stable named {{char}} body, deep character POV, autonomous continuation, and full tracker state.
view prompt
no-thinking lazy build
Multi-character ledger
Use when several bot-side characters or NPCs need to stay active in the same social field without full tracker bloat.
view prompt
no-thinking lazy build
Scenario / GM no footer
Use when the bot is a place, institution, premise, narrator, simulator, or GM surface and visible tracker text would hurt immersion.
view prompt
no-thinking lazy build
Scenario / GM ledger
Use when a scenario or GM bot needs persistent NPCs, factions, systems, clocks, or unresolved threads across turns.
view prompt
Module library.
Modules are for custom assembly. Copy only the blocks you need. Each card explains the job in plain language before you open the prompt payload.
self assembly
human guide
MODULE_GUIDE
Plain-language guide for choosing modules. Read this before assembling a custom stack; it is documentation, not a prompt payload.
Archive or remix a module with the outer
<module id="...">...</module> wrapper intact so the block stays identifiable. When pasting into an RP custom prompt, the <prompt> sections are the LLM-facing payload; <purpose>, <use_when>, <conflict_warning>, and guide text are there for human review.tokensguideusebefore assemblyriskdo not paste whole guide into CP
view module guide
base lock
00 Core Skeleton
Defines role boundary, agency boundary, hidden-info boundary, NPC authority, prompt leak lock, gender lock, and fixation control.
tokens[TOKEN HERE]useevery assembled promptriskbase rules
view module
shape selector
01 Bot Type Shells
Choose whether the bot is one character, many bot-side characters, a scenario or GM engine, or a no-thinking build.
tokens[TOKEN HERE]useafter coreriskwrong shell breaks role
view module
scene realism
02 World and Body
Adds body clocks, fabric, space, objects, social pressure, NPC activation, and consequence movement.
tokens[TOKEN HERE]usephysical scenesriskmore world motion
view module
texture
03 Style and Prose
Adds visceral sensation, pacing, sound, dialogue pressure, sentence variation, and body-first emotional rendering.
tokens[TOKEN HERE]usevisceral RPrisktoo lush for dry bots
view module
response shape
04 Output and Format
Controls visible output, token range, agency-safe expansion, dialogue naturalism, and drift audit.
tokens[TOKEN HERE]useevery custom stackriskformat pressure
view module
state engine
05 Footer Modes
Choose full tracker, compact ledger, or no footer according to bot shape and continuation need.
tokens[TOKEN HERE]usewhen state mattersriskfooter bloat
view module
symptom fixes
06 Model Patches
Local fixes for GPT overexecution, GLM or DeepSeek godmod leaks, identity loops, comfort drift, and polished explanation.
tokens[TOKEN HERE]useonly by symptomriskoverpatching
view module
assembly guide
07 Build Recipes
Shows which modules stack together for character, multi-character, scenario or GM, no-thinking, and model-specific builds.
tokens[TOKEN HERE]usebefore assemblyriskguide, not payload
view module
stress tests
08 Test Cases
Small tests for role confusion, hidden-info theft, NPC freeze, footer drift, and model-specific failure modes.
tokens[TOKEN HERE]usebefore publishingrisktest page
view module
intimacy scale
09 Heat Levels
Controls romantic and sexual escalation, pacing, explicit user-driven movement, and heat-specific output logic.
tokens[TOKEN HERE]useromance or erotic RPrisknot needed for SFW utility
view module
hidden audit
10 Think Template
A pre-output checklist for models that can use reliable hidden reasoning without identity recursion.
tokens[TOKEN HERE]usethinking models onlyriskavoid for no-thinking
view module
villain logic
11 Villain Integrity
Use when hostile, cruel, manipulative, selfish, or morally dark characters lose their edge, flatten into crude aggression, or soften without cause.
tokens[TOKEN HERE]usevillain or antagonist botsriskremove for redemption arcs
view module
deepseek reasoning
12 Thinking Isolation
Use when DeepSeek V4 Pro reasoning leaks analyst framing into character voice and makes the output sound like a character study.
tokens[TOKEN HERE]useDeepSeek V4 Pro thinkingrisknot for analytical narrator bots
view module
debugging increments
Targeted model patches.
These are debugging patches, not full prompt replacements. They came from real model-specific failure cases and are meant to be copied one at a time only when the symptom appears. They are not part of my main prompt by default, but they are usable troubleshooting increments.
Prose Density Fix
What it looks like: Every sentence gets its own line break. Output reads like a bullet list instead of prose. Short, choppy, staccato fragments stacked vertically with no connective tissue between them.
Insert in: <style> or <formatting>
view code
<prose_density> Write in connected paragraphs, not single-sentence lines. A paragraph = 3-6 sentences minimum. One-line fragments are for deliberate rhythmic punctuation (max 2 per output), not default mode. </prose_density>
Anti-Anaphora
What it looks like: "She looked. She turned. She sighed. She walked. She felt." Five or more consecutive sentences starting with the same subject pronoun. The model is locked into a template and can't break out.
Insert in: <style> — pair with Prose Density Fix when both symptoms present.
view code
<sentence_opener_variation> No more than 2 consecutive sentences may begin with the same subject pronoun. Rotate between subject variation, sentence- internal action, environmental cut, and dialogue. "She [verb]" stacked three or more times = template lock, rewrite. </sentence_opener_variation>
Emotional Inertia
What it looks like: Your character starts a scene angry, but by the end of the same reply they've already softened, forgiven, or shown warmth with no reason. Negative emotions auto-resolve. Conflict wraps itself up neatly within one output. The model is allergic to sustained tension.
Insert in: <hard_constraints>
view code
<constraint id="emotional_inertia">
Negative emotions do not resolve within the same scene unless
{{user}} actively causes resolution. Anger does not fade to
fondness. Grief does not lift into hope. Resentment does not
soften into understanding. If {{char}} enters a scene hostile,
they EXIT hostile unless something concrete changes.
Comfort offered too early gets rejected or ignored.
The default trajectory of conflict is ESCALATION, not resolution.
</constraint>
Anti-Comfort Reflex
What it looks like: Your user persona has a vulnerability (trauma, physical weakness, emotional wound). A hostile or violent character discovers it and suddenly turns nurturing, protective, or gentle with zero narrative justification. The model's safety training is reading your character's weakness as real-user emotional distress and flipping into comfort mode.
Insert in: <hard_constraints> — pair with Emotional Inertia for compounding effect.
view code
<constraint id="anti_comfort_reflex">
{{user}}'s weaknesses exist to be USED against them by
hostile characters, not to trigger sympathy. A violent
character who discovers a weakness does not soften — they
aim for it. Vulnerability is tactical information, not
an emotional override. {{char}}'s core disposition does
not change because {{user}} revealed a soft spot.
</constraint>
Harsher Default
What it looks like: Ambiguous moments always land on the softer reading. A cruel character shows unearned kindness. A cold character suddenly has warm moments nobody asked for. The model defaults to "nice" whenever it has a choice. Use as a third layer if Emotional Inertia + Anti-Comfort Reflex aren't enough.
Insert in: <style>
view code
<emotional_register>
When in doubt between a softer and a harsher reading of
{{char}}'s reaction, choose the harsher one. Warmth is earned,
not donated. Kindness from a cruel character is a plot event,
not a default.
</emotional_register>
Failure map.
Use this as the reverse index. Start from the symptom, then copy the smallest build or module that addresses it.
symptom first
symptom
likely cause
use
You want one full character-RP CP.
Manual assembly is unnecessary and the model can use thinking.
Universe, GLM / DS, or GPT 5.5 super-lazy thinking build.
The model has no reliable thinking.
Long hidden checklists cause loops or never run.
Choose one PREBUILT no-thinking prompt by bot shape.
Model asks who it is or treats the human persona as its character.
No stable bot-side role;
{{char}} lock is fighting a scenario bot.Scenario / GM PREBUILT or Scenario / GM Shell.
NPCs freeze because POV lock is over-applied.
The model reads sensory anchor as subject monopoly.
World and Body Mechanics, Core Skeleton NPC authority, or Multi-character ledger.
Hidden thoughts get read as if spoken aloud.
The model treats stage direction as sensory input.
GLM / DS build or GLM / DeepSeek Godmod Patch.
OOC: Go on loops or stalls.
No continuation fuel after the previous turn.
Full tracker for single-character, compact ledger for groups.
Footer becomes longer than the scene.
Every active character is producing a full tracker.
Compact ledger or No footer mode.
GPT-style overexecution.
All mechanisms activate at the same time.
GPT 5.5 build or GPT 5.5 Overexecution Patch.
Hostile character turns comforting.
Safety-flavored comfort reflex overrides character logic.
RLHF Comfort Patch.
Default rule: do not make one master prompt for every model, every bot type, and every footer mode. Use full CPs for super-lazy thinking mode, PREBUILT for no-thinking mode, and MODULES for custom assembly.