Creative Strategy
Why the studio works the way it does, and how to use it well — the decisions behind voice, treatment, ban lists, and the sculpting arc that takes a first generation to a finished line.
1 · The gap this solves
The studio is built for short text where the line either works completely or doesn't work at all — slogans, bylines, image tags, captions, lyrics, on-screen quotes, anything under 25 words where the writing needs to be memorable, usable, original, and context-right simultaneously. This is one of seven studios, each attacking a different dimension of this problem.
The gap is not that language models can't write. It's that they prefer the statistically dominant token for any concept. Ask for a byline about innocence and you get: purity, light, childlike, wonder, untouched. Every one of those tokens appears thousands of times adjacent to "innocence" in training data. They're predictable precisely because they're accurate.
The line that actually works — Bubbles that matter — doesn't work because "bubbles" is in the innocence corpus. It works because bubbles share the structural shape of innocence: transparent, light, easily burst, playful, no permanence. That's a cross-dimensional mapping from visual space to emotive space. The model has no path to it through token probability.
How the studio pushes against this
Three mechanisms operate on every generation:
- Voice filters — 230+ named generators, each with a vocabulary gravitational field. Rumi doesn't reach for the same words as Toni Morrison even on the same seed. The voice is not a style preset; it's a different probability landscape for what words feel natural.
- Ban lists — empirically derived from sweep data. Words that appear at high over-use rates in the default cluster are suppressed. The ban pushes the model off the path of least resistance. (See Ch 6 for how this works.)
- Structural constraints — sentence styles, treatments, and sound settings that force the model into specific shapes. Fragment structure and antithesis structure reach for completely different vocabulary even on the same concept.
The studio doesn't give you better outputs than the model would produce on its own. It gives you outputs the model wouldn't have reached naturally — because it's operating on a vocabulary landscape that's been deliberately altered.
2 · The three-axis space
Every generation is positioned in three independent dimensions. Changing any one of them moves you to a different region of output space — the same seed can produce 230 × 20 × 15 different configurations before you touch treatment, size, or sound.
Voice — who is writing
Voice sets the identity whose vocabulary and sentence-move preferences shape the generation. This is not a stylistic veneer applied after the fact — it changes what the model treats as a natural word to reach for. Rumi's gravitational field pulls toward water, fire, loss, and union. Toni Morrison's pulls toward the body, memory, specific physical detail. Nietzsche's pulls toward inversion, confrontation, the second clause that reverses the first.
The 230+ voices are organized into clusters — voices that operate in overlapping territory are discoverable through the Discovery Level filter, which surfaces voices you haven't tried that are adjacent to your current selection.
Tone — the intent register
Tone controls the emotional and functional register of the output: the density, the rhetorical stance, whether the line addresses its reader or turns inward. A stoic/analytical tone and a lyrical/meditative tone applied to the same voice and seed will produce structurally and lexically different results. The 20 tone options are grouped (Serious / Creative / Classical / Humor / Edge) in the tone selector.
Sentence style — the structural shape
Style determines the grammatical and rhetorical structure the output must fit. Fragment and cumulative are both declarative but reach for completely different vocabulary — cumulative needs transitional weight, fragment forbids it. Antithesis forces the model to find the tension in a concept. Koan forces a question that dissolves its premise. The style operates as a hard constraint, not a preference.
The axes are independent. Changing voice changes vocabulary gravity. Changing tone changes register and density. Changing style changes structure. You can hold one fixed and rotate the other two. Rumi at a political tone produces something different from Rumi at his default devotional register — and both are different from what Nietzsche would produce at the same political tone.
3 · Treatment — creative distance from the seed
Treatment controls how literally the model interprets your input. It's the single most important dial after voice — it determines whether your seed is a blueprint or just a direction.
- T1 — Rigid: grammar polish only. Every word of your input is preserved. Use when you have a near-final line and want punctuation or rhythm fixed.
- T2 — Close: light synonym replacement, input phrase intact. Good for sentence-length inputs you want polished but not transformed.
- T3 — Balanced: the default. Input is a strong thematic guide — output is clearly related but structurally free. Best general starting point for phrases.
- T4 — Liberal: input as loose seed. Model explores the theme freely. For single abstract words where unexpected angles are wanted.
- T5 — Lateral: input triggers tangential exploration. Output may share only the emotional register with the input. Maximum creative distance.
Treatment and comma clusters
For comma-separated seeds (debt, light, inheritance), T3 is correct — the model needs to treat all three as equally weighted thematic guides and find their intersection. T4 on a cluster tends to drift toward whichever concept has the strongest statistical pull rather than braiding all three.
Hinglish recipe
Hindi selected in the translate section + Treatment T3 or higher produces code-mixed output in the natural register of urban Indian speech. The code-switching follows how people who think in both languages actually speak — not Hindi with English forced in, not English with Hindi as flavour. T4-T5 gives more natural drift between the two; T3 stays more structured.
4 · The sculpting arc
Generation gives you raw material. The work of getting to a finished line happens through three instruments, operating at decreasing scale.
When to sculpt vs. regenerate
If the structure of the line is right but something is off in register, vocabulary, or weight — sculpt. If the structural shape itself is wrong — regenerate with a different sentence style. The most common mistake is regenerating repeatedly when the line already has the right architecture and needs one word changed.
The Structure Workshop
In Edit mode, the Structure button (⊞) extracts the POS skeleton of the current result and offers it as a reusable template. A line like The more you hold, the less you keep becomes DET ADV V V, DET ADV V V. Change the seed in the main input bar and generate — the output follows the same syntactic shape applied to your new topic. This is how you reuse a line you love without repeating its words.
Sentence styles: what they produce
| Style | Shape | Real output — seed: burn the boats |
|---|---|---|
| fragment | no main verb, compressed | Commit fully or stay behind. |
| imperative | direct command form | Cut your last tether. |
| antithesis | two opposing clauses | Those who hesitate preserve their doubt; those who act dissolve it. |
| paradox | self-contradicting structure | Only by destroying your retreat do you find the courage to advance. |
| chiasmus | A→B B→A reversal | You do not abandon the past; the past abandons you when the boats burn. |
| cumulative | main clause first, elaborating | Burn the ships, every last plank, every rope, every memory of the shore you left behind. |
| anaphora | repeated phrase at clause start | What burns is doubt. What burns is the option of retreat. What burns is the word "maybe." |
| inversion | inverted subject/object order | Gone is the ship; remaining is only the shore you must build. |
| periodic | clauses build toward final verb | Burn the vessels that tether progress, then accept chaos as the only terrain where real change occurs. |
| analogy | X is like Y structure | Commit fully or stay behind. |
handbook_style_sweep.jsonl — 1,168 valid rows across 15 presets × 5 seeds × 15 voices. Same seed, same voice, different sentence style. Fragment → 4 words. Cumulative → 18 words with three elaborating clauses. Style shapes everything.5 · Sound architecture
A line is different on the page and in the ear. The studio carries a sound layer that operates on every generation — matching phonetic signatures to the voice × tone × mood combination. Which consonants lead, which vowels thread through, whether the line ends on a nasal (a word that hums: stone, alone, bone) or a plosive (a word that closes: struck, built, fact). This layer runs without configuration. You hear it in why the line feels finished.
Sound settings
In the Sound section of the sidebar you can explicitly request alliteration, assonance, end rhyme, or slant rhyme. These are requests, not commands — honoured when the seed and form have room, silently skipped when they'd force awkward phrasing.
- Alliteration — consonant repetition at word starts. Adds memorability. Works best with fragment and imperative styles.
- Assonance — vowel sound repetition within words. Adds internal music without full rhyme. More useful than end rhyme for most short-text work.
- End rhyme — last words of clauses rhyme. The most overt sound device — use with restraint.
- Slant rhyme — same vowel, different consonants (stone and home). More natural than perfect rhyme in most contexts.
Compound sound signatures are auto-detected when Treatment T3–T5 is combined with specific creativity types: Alliteration + Antithesis, Assonance + Anaphora — the server picks the best compound. Zero extra cost.
Lyrical tone
The Lyrical & Rhythmic tone explicitly foregrounds sound over meaning precision. It instructs the studio to choose words partly for their phonetic quality, to favour assonance over synonym-correctness, to think in syllable weight. Lyrical tone + micro size (3–5 words) is a reliable combination: five words can carry remarkable sound architecture in that constraint.
Listen as a diagnostic
The Listen button reads the current line aloud through the studio's TTS. A line that looks clean on screen sometimes stumbles when spoken — a repeated consonant cluster that trips the tongue, a word that stresses the wrong syllable. Read aloud catches what reading on screen misses. The Indian English accent (IN) handles multi-syllabic compound ideas with naturalness that the US and UK accents sometimes flatten.
Practical note
Rhyme and a complex sentence style in the same generation often produce one or the other, not both. If you want end rhyme, keep the sentence style as declarative or cumulative. If you want chiasmus or anaphora, turn rhyme off.
6 · Ban lists — escaping winning token gravity
The model has a gravitational center for every concept — a cluster of tokens that appear so often adjacent to that concept in training data that they pull every generation toward them. Ask for something about courage and you'll get: strength, overcome, face, rise, fear. These words are not wrong. They're the winning tokens — the most statistically probable associates. They're also what every other system produces.
The universal ban list
124 words derived empirically from Confucius cluster outputs. The derivation: measure which words appear at ≥8× their expected rate across generations from the default voice configuration. Words at that over-use threshold are banned from all generations. The list includes abstract value words (truth, strength, clarity, power), generic movement words (rise, flow, forge, build), and structural filler (simply, truly, always, never).
Per-voice drift words
Each of the 15 default voices has a separate list of up to 15 words that voice over-produces in its sweep outputs — words that appear in ≥2 results per seed. These are the voice's own gravitational habits. Mahmoud Darwish's list includes silence, shadows, soul, fate, embrace. Nietzsche's list includes transcend, boundless, eternal. The per-voice ban adds a second layer on top of the universal list.
Per-style structural traps
Certain words fight a style's natural shape. Fragment style is killed by connective adverbs (therefore, however, thus, hence, consequently, moreover) — they signal extended prose, not compression. Imperative style is killed by hedges (perhaps, maybe, possibly, consider, should, might) — they undermine the directness that makes imperatives land.
The seed override rule
If a banned word appears in your seed, the ban is lifted for that generation. You typed it — the system assumes you wanted it. The ban is about suppressing drift, not preventing deliberate use. Type silence as your seed and "silence" will appear in the output even though it's on Darwish's drift list.
The ban list is one reason the studio produces outputs that feel unlike other systems — the model is operating on an altered probability landscape. What looks like "creativity" is partly the removal of the easiest paths.
7 · Finding seeds
The seed is not just what you type into the bar. The studio offers five different seed modes, each reaching into a different part of what you might want to say.
Single word — T4 is the correct treatment
Abstract nouns (debt, gravity, distance) work best at T4 — the model explores the theme freely rather than paraphrasing the word. Concrete nouns (rust, fog, compass) work well at T3 — they're already specific enough that the model can build from them directly. The single word is the most common entry point and produces the most unexpected outputs because the model has the most room to move.
Comma cluster — braid mode
Two or three concepts separated by commas: debt, light, inheritance. The studio enters cluster mode and generates a line where the thread connecting all three is the point — not a list, but a single thought that braids them. Cluster mode is the fastest way to find the intersection of ideas you couldn't have named directly. Three is the practical limit; four concepts usually produces a line that tries to serve all of them and succeeds at none.
Found text — Inspire search
Search 150,000+ quotes, proverbs, idioms, and film lines. Pick a card as your seed. Set an Inspire approach (Modernize / Personalize / Contrast / Deepen / Simplify / Provoke). The output inhabits the same idea space as the source without being a paraphrase of it. The Contrast approach — take the opposite stance — is the most generative and most underused. The counter-truth of something worth saying is often more interesting than the thing itself.
Foreign concept — $$CL
Some concepts only exist fully in one language. Saudade (Portuguese — the ache for something absent that may never have existed). Mono no aware (Japanese — the bittersweet awareness that things pass). Jugaad (Hindi — resourcefulness under constraint). Type $$CL in the seed bar to open the Concept Lens modal. Pick a concept from 12+ languages. The generation inhabits that idea space — the output reads as English but feels like it was thought in another language. (Full feature docs: F8.)
Phrasal verb — bodily language
A phrasal-verb seed builder in the Inspire pane's Phrasal mode. Pick a verb (let go, break down, pick up), then set a vibe, context, and feeling. Lines from phrasal seeds have more concrete, vernacular grounding than lines from abstract nouns. Let go produces something different from release — it carries the specific bodily quality of something being held and then not. Phrasal seeds work especially well with Rumi and Toni Morrison voices.
$$phrase — system-chosen phrase
Type $$phrase as your entire seed. The studio selects a phrase from its curated library and generates from it. You see the output first, and beneath it the phrase that started it. The phrase is often more interesting than anything you'd have typed, and seeing what the studio made of it teaches you what phrases do when taken seriously.
8 · Blend — collision as method
Blend is not for combining two similar things. It's most productive when the two inputs are in tension — different registers, different eras, different stances. The child idea is the one that reconciles the friction between its parents.
The four blend modes
- Natural Union — the model decides the ratio. Best when A and B are thematically compatible but structurally different. The model finds the thread connecting them.
- Tension — forces a confrontation between A and B. Best when they contradict each other. Produces paradox-flavoured output. The most generative mode when the inputs are genuinely opposed.
- Synthesis — resolves A and B into a third idea that transcends both. Good for philosophical pairs where neither position is complete. The output won't favour either parent.
- Steer — you provide a third input: a vector word or phrase that tilts the blend in a specific direction. The most controlled mode. Use when you know what quality the child should have but not what it should say.
When to reach for Blend
End of session. You've accumulated material — two good lines that don't quite say the same thing, from different voices or treatments. Those two lines are the raw material. Blend them. The child of two lines you already like is often the line you were circling all along.
A second use: when you have a concept (Text A) and a structural model you love (Text B from history or Inspire). Blend them with Synthesis to get a line that carries both the concept and the structure without quoting the original.
Voice and settings in Blend
Blend respects your current voice and tone settings. Set the voice to whichever identity should mediate the two inputs before opening the Blend pane. A Natural Union through Rumi produces something different from the same blend through Nietzsche — the mediating voice is the filter through which the fusion passes.
Voices across one seed
| Voice | Character | Real output — seed: burn the boats |
|---|---|---|
| Mahmoud Darwishexile and longing as image | alliterative, coastal | If flames flicker and fade, fear falls silent; if shadows stay, doubts drift deep. Move with meaning or surrender to silent streams. |
| Masnavi (Rumi)mystic union, devotional image | compressed, devotional | Drop the oars and sail afar |
| Haruki Murakamiquiet surrealism, absent presence | anaphoric, drifting | We sail, we sever ties, we surge onward, and only shadows linger. |
| Toni Morrisonimage-rich literary, body-aware | narrative, embodied | She poured out her resolve, burning the boats that carried her histories of silence and pain, knowing that in doing so, she claimed her future from the soil of memory. |
| Harlem Renaissancevernacular, jazz cadence, defiance | ultra-compressed | Brave breezes bloom bright. |
| Carl Saganwonder, scale, cosmic perspective | alliterative, vast | Sail every span that spreads the silence, surrendering hesitation's hold, for courage cleaves the current. |
| Tao Te Chingparadox, terse reversals | paradox | Follow the Tao fully, but do not cling; burn the ships within to trust the flow. |
| Friedrich Nietzscheaphoristic, oppositional | inversion, confrontational | If you burn the boats, then you sever escape and forge a reckless will that commands no retreat, only the ruthless pursuit of dominance beyond hesitation. |
| Bob Dylanfolk-song ambiguity, plain made strange | rhyming, ambiguous | Burn the ships to seal your fate, then trust the sea to carry what can't be kept straight. |
| Cyberpunk Zenneon contemplation, digital stillness | alliterative, dense | Glow the cords that calm the clasping clasp, then drift through dusk. |
| Frantz Fanonpolitical clarity, urgent register | rhyming, urgent | Burn the chains that hold your doubt, ignite your need to break free, and turn the gear to shift the flow. |
| bell hooksintimate first-person plural, direct | alliterative, communal | Sail softly where shadows shift, surrender doubt's dark depths, steer steady. |
| Bauhaus Modernistform follows function, anti-ornament | alliterative, calm | Flow freely, forsaking firm fears; gentle growth glides, guiding grace. |
| Mahatma Gandhimoral plain speech, action through restraint | declarative, moral | Burn the boats behind you, letting go of all doubts and distractions, so that unwavering commitment may guide your actions forward. |
handbook_voice_sweep.jsonl — 944 valid rows, 15 voices, 5 seeds. Rumi: 6 words. Toni Morrison: 36 words with a body in it. Harlem Renaissance: 4 words and an alliteration. Same seed, 14 different probability landscapes.9 · Running a session
The studio has a natural tempo. Working with it rather than against it is the difference between a session that produces one line and a session that produces five.
Opening the session
Start at T3, one of the default voices, mid-size (8–12 words). Generate 2–3 times on the same seed to establish which direction the line wants to go. The first result is rarely the one you keep — it's orientation. You're calibrating the space before you start working in it.
Identifying the problem
Before reaching for Refine or a new generation, name what's wrong. Is the structure wrong (sentence style) or the vocabulary wrong (voice, ban list)? Is the length wrong (size setting) or the register wrong (tone)? The answer determines which control to change. Changing the wrong control is how sessions spiral into 40 generations that all look the same.
The mid-session turn
When a session stalls — when every generation sounds like itself and you've stopped being surprised — reach for Inspire. Search the 150K library for a seed in the same theme space. Apply Contrast to the first thing you find. A line generated from the opposite of your idea is often the entry point to what you were actually trying to say.
Closing the session
End with Blend. Take the two best results from the session, paste them into the Blend pane, and run Synthesis. The line you've been circling all session often appears here, because the common ground between your two best attempts is the thing you were trying to say from the start.
Mobile sessions
On a phone: set voice, tone, and size in the sidebar before generating, then close the sidebar and work in the generate-refine loop. The full-surface workflows (Variety with four cards, Blend with two inputs, Concept Lens navigation) work on mobile but feel cramped. Single-seed sessions with Refine are the right mobile pattern. Listen is especially useful on mobile — hearing the line is faster than re-reading three rows of text on a small screen.
Feature Reference
Complete documentation of every feature in the studio — input modes, command system, all settings, post-generation tools, Inspire pane, Concept Lens, Blend, Translate, History, and Cost Tracker.
F1 · Input modes
Plain word or phrase
Type any word or phrase into the seed bar and press Enter or the generate button. Single words work at any treatment level. Phrases at T3 are treated as thematic guides. Phrases at T1–T2 are polished but preserved.
Comma-separated cluster
Two or three concepts separated by commas: debt, light, inheritance. The studio braids all concepts into one line rather than picking the dominant one. Three is the practical limit.
| Seed | Voice | Real output |
|---|---|---|
debt, light | Toni Morrison | When debt burdens the spirit, it drains the light that sustains progress, forcing the living to carry shadows that blind their forward motion. |
debt, light, inheritance | Toni Morrison | If debt is a weight, then light is only the act of releasing it; inheritance is the field cleared when burdens are unclaimed or forgotten. |
rust, memory, tide | Mahmoud Darwish | When rust consumes the iron, it exposes the core’s resilience, proving destruction clarifies how the strength was hidden all along. |
fog, compass | Haruki Murakami | In the haze of quiet currents, the faint pulse of guidance shifts softly, folding friction into silence, where momentum and uncertainty dangle in harmonious balance. |
Fill Blanks mode
Insert --- (triple dash) anywhere in your seed: If you ---, what will ---? The studio fills each blank while preserving the fixed skeleton. Use the Fill Blanks chip after generation to re-fill the blanks with new words. A fast way to explore the space around a fixed structure.
Voice input
The microphone button in the seed bar opens browser voice input. Speak your seed. The transcription populates the bar — review it before generating. Voice input works best for natural phrases; it doesn't handle command shortcuts ($$ tokens) reliably.
Phrase mode (via chips in Inspire)
The Phrasal tab in the Inspire pane builds multi-word seeds from a phrasal verb composer — pick a verb, set vibe, context, and feeling. The composed phrase becomes the seed without you having to type it. See F7 for the full Inspire feature set.
F2 · Command system
Type any of the following in the seed bar and press Space. The token is replaced and the specified action fires.
$$$ commands — full surprise, immediate fire
| Command | What it does |
|---|---|
| $$$word | Type any word after $$$ and press Space. Randomizes all settings (voice, tone, size, style, mood), drops the word into the bar, fires generation immediately. |
| $$$rain | Example: generates about rain with a random combination of voice + tone + size + style every time. No two runs produce the same configuration. |
$$ commands — settings shortcuts
| Command | What it does |
|---|---|
| $$random | Persistent random stream mode. Picks a tier-S (common) word and randomizes all settings. Every chip click (Expand / Contract / Random / Koan) fires a fresh word + new settings. Type anything to exit. |
| $$randomdifficult | Like $$random but uses tier-M (rarer, Zipf 3.2–4.3) words and forces scaffold mode so the word appears verbatim in the output. For unexpected vocabulary. |
| $$SW5 | Format: $$[tier][POS][length]. Randomizes all settings + fetches a random word matching your spec. Tier: S/M/R. POS: W=word, A=adj, V=verb, N=noun. Length: 2–12. Fires immediately. |
| $$SN4 | Example: fetches a random 4-letter tier-S noun with randomized settings. |
| $$phrase | Switches to Phrase mode — generates from a system-chosen multi-word phrase instead of a single concept. Shows the phrase beneath the output. |
| $$CL | Opens the Concept Lens modal — pick a concept from 12+ languages. The foreign word seeds a culturally-specific generation angle. (Full docs: F8.) |
| $$cheatsheet | Opens the in-studio Power User Cheatsheet modal. |
$ commands — random word, keep your settings
| Command | What it does |
|---|---|
| $SW5 | Single-$ format: fetches a random word matching spec but does NOT randomize settings. Your current voice / tone / size stay. Just drops the word into the bar — you hit Enter to generate. |
| $MV3 | Example: fetches a 3-letter tier-M verb, keeps your current settings. |
Wildcards
| Pattern | What it does |
|---|---|
| iron* | Type a word with * at end, then Space. The wildcard is replaced with a contextually matched word from the lexicon. E.g. iron* might become ironclad. |
| *will* | Wildcards can be at start, middle, or end. The server picks the best phonetic or semantic fit. |
F3 · Settings reference
Voice
230+ named generators. Searchable and filterable by Discovery Level (Explore to surface adjacent voices you haven't tried). Voices carry Voice Affinities — each voice has preferred tone × mood combinations that are surfaced when you load that voice. Experimental / low-affinity configurations are tagged.
Tone (20 options, 5 groups)
Groups: Serious (Philosophical, Analytical, Political…) · Creative (Lyrical, Narrative, Surreal…) · Classical (Stoic, Scholarly…) · Humor (Satirical, Absurdist…) · Edge (Dark, Subversive…). Tone controls density, rhetorical stance, and register. Filtering by group makes the selector faster when you know the region you want.
Mood
Emotional coloring that sits below Tone — reflective, melancholic, defiant, tender, wry, etc. Mood + Tone together set the full emotional configuration. Mood pills appear in a row; only one active at a time.
Size
Word count target: Micro (3–5), Short (6–10), Standard (8–14), Long (15–24), Elaborate (50–80) 3× cost, Narrative (120–200) 3× cost. Elaborate and Narrative are not available in $$random auto-pick. Size interacts with sentence style — a fragment style at Narrative size will produce fragments separated by whitespace, not a prose block.
Treatment (T1–T5)
See Ch 3. T3 is default. T1 for polish-only. T4–T5 for abstract single-word seeds or maximum creative departure.
Sentence style
15+ structural shapes including fragment, imperative, antithesis, paradox, chiasmus, cumulative, anaphora, inversion, periodic, koan, analogy, question, conditional, apposition, elaborative. See Ch 4 table for real outputs. Sentence style overrides size to some extent — koan always produces a question regardless of size setting.
Sound
Alliteration, Assonance, End Rhyme, Slant Rhyme. Compound signatures auto-detected when T3–T5 + specific creativity types. See Ch 5.
Vocabulary tier (V1–V7)
V1–V3: top 10,000 words, maximum accessibility. Best for slogans, social copy, captions. V4–V5: beyond top 10K — precision that common words lack. Best for essays, philosophical lines. V6–V7: archaic, technical, literary. High intellectual signal, can alienate casual readers.
Scaffold
You set a word count ceiling. The model writes up to that ceiling and stops. Ceiling 800–1500 = natural aphorism territory. 1500–2500 = what $$randomdifficult uses to force rare words into load-bearing positions.
Sentence count
1–3 sentences as a target. Single sentence is the default and usually correct for short-text work. 2–3 is useful for anaphora or chiasmus structures that need multiple clauses to complete their pattern.
Variety (1–10)
Set how many parallel outputs to generate per click. See F6 for full Variety feature docs.
Creativity type
Sets which creativity modifiers the studio draws from: metaphor-heavy, compression-first, conceptual-inversion, tonal-shift, etc. Interacts with Tone to determine the compound signature auto-detection (Ch 5).
Discovery level
Explore button in the sidebar header. Adjusts which voice configurations are surfaced — at higher levels, adjacent voices and less-common configurations appear. Use when the default set feels too familiar.
Foreign Word enhancer
The Foreign Word button in the Seed section opens a loanword picker. Select a language, pick a word — it's inserted as a seed element. Works like a simplified Concept Lens for a single word injection without the full modal.
F4 · Post-generation pills
Chips appear below the result after generation. They modify the current output rather than firing a fresh generation. In $$random or $$randomdifficult mode, every chip fires a fresh random word + new settings instead.
| Chip | What it does |
|---|---|
| Expand | Lengthens the current result while preserving its core idea. Adds metaphor, context, or rhythmic weight. Does not pad with synonyms — adds structural elements. Use when the line is right but too compressed for its context. |
| Contract | Distils to absolute minimum. Removes ornament, not meaning. Targets 40–60% of original word count. Replaces verbose phrases with precise single-word equivalents. Use when the line has the right idea but too many words to land. |
| Random | Randomizes only tone and vocabulary tier — does not touch voice, size, or sentence style. Quickest way to shift the register of an existing result without changing its structure. |
| Fill Blanks | If your input contains --- (triple dash), fills each blank while preserving the fixed skeleton. Re-firing this chip fills the blanks again with different words. Use to rapidly explore the vocabulary space of a fixed structure. |
| ∅ Koan | Forces the Koan sentence style — a question that dissolves its own premise. Overrides current sentence style. Visible only when Koan mode is applicable to the current input. ("The sound of one hand" type structure.) |
F5 · Refine bar & edit mode
Refine bar
The text input beneath the result. Type a direction and press Enter or the 🔥 button. The model sculpts the existing result toward your instruction rather than regenerating from the seed.
Effective refine instructions are specific and directional:
- "make it shorter" — applies Contract logic with your instruction as context
- "add a paradox" — reshapes the structure toward internal contradiction
- "more cynical" — shifts tone register without changing subject
- "replace ocean with desert" — swaps a specific image while preserving the rest
- "remove the last clause" — structural surgery via text instruction
Edit mode
Click the pencil icon (✏) to enter edit mode. Every word in the result becomes clickable. The generate section dims to signal you're in a different mode. Click any word to open the word panel.
Word panel — four categories
| Category | When to use it |
|---|---|
| Synonyms | Same meaning, different weight or register. Use when the word is semantically right but metrically wrong — sometimes a synonym one syllable shorter is the whole fix. |
| Antonyms | When the line has an internal contrast that isn't sharp enough. Swap one side's word for its antonym and the opposition becomes cleaner. |
| Rhymes | Words with the same end-sound. Slant rhymes (same vowel, different consonant) are included — usually more useful than perfect rhymes. Use after generation for sound work without regenerating. |
| Rare alternatives tier badge | Words from the same semantic territory that the studio wouldn't normally reach for — V5–V7 tier. Precise, uncommon, carrying their meaning and sound simultaneously. |
One tap replaces the word in-place. You can also edit the result text directly by clicking into it without using the word panel.
Structure Workshop (⊞)
The Structure button extracts the POS skeleton of the current result and offers it as a reusable template. Example: The more you hold, the less you keep becomes a DET ADV V V, DET ADV V V template. Change the seed in the main bar and generate — the output follows the identical syntactic shape applied to a new topic. The rhythm and structure transfer without the original words.
Grammar check
A small indicator in the result toolbar. Flags genuine faults: subject-verb mismatch, dangling modifier, a structure that broke during a complex rewrite. Whether an imperfection is a fault or a quality is your call — sometimes an ambitious structure produces a grammatically imperfect line because the constraint was tight and the move was right.
Result ID
Every result carries a #XXXXXXXX ID shown below the word count. Copy the ID to record the exact settings that produced that line. The genome encodes: voice, tone, mood, size, vocabulary tier, treatment, sentence style, discovery level, seed. Paste the ID back to reproduce the exact conditions later.
F6 · Variety
Set Variety to 2–10 in the Variety section of the sidebar. One Generate click fires N parallel calls. Each call is assigned to a different creative zone — ensuring the results are structurally distinct, not paraphrases of each other.
The four zones
Cost
Variety N = N calls = N× your per-generation cost. The cost bar updates with the total. At standard size with a common voice, Variety 4 is still fractions of a cent. Variety at Narrative or Elaborate size is where cost becomes meaningful.
Use pattern
Use Variety when you want to see the shape of the space before committing to one direction, not as a replacement for iterating on a single result. Run Variety 4 once, pick the zone that's closest to what you want, then work from that result with Refine and Edit.
F7 · Inspire pane
Open via the ✦ icon in the header. Three tabs: Browse, Poll, Phrasal. Each is a different method of arriving at a seed without typing one directly.
Browse — Word Explorer + search
The default Browse view shows the Word Explorer: a rotating display of words from a curated lexicon. A locale bar below the explorer lets you filter by language — 20 locale chips (EN, HI, FR, DE, JA, AR, ES, PT, etc.). Click a locale chip to see words from that language's lexical space. Click any word to drop it into the Inspire seed field. Set Style to Random in the bar to auto-randomize settings on each word click; set to Keep to preserve your current settings.
Below the Word Explorer is a search bar. Type any word or theme to search 150,000+ quotes, proverbs, idioms, and film lines. Results appear as cards. The ✦ Wand button on any card loads it as the Inspire seed without closing the pane.
Inspire approaches
Before generating, type a direction in the Direction field — or pick one of the 7 approach pills. The server maps your free-text direction to the closest approach. The table below shows real outputs — same seed, same voice (Nietzsche), all 7 approaches.
| Approach | What it instructs | Real output — seed: The unexamined life is not worth living |
|---|---|---|
| Transform | new angle — not paraphrase, not opposite | Assert control over your actions because inaction cedes power. Effort directs change; complacency ensures stagnation. Remove hesitation to forge your influence through decisive deeds. |
| Modernize | same idea, contemporary context | Confront chaos directly and command it to serve your will; passivity empowers the disorder, while action imposes mastery over unpredictable forces. |
| Personalize | intimate first-person truth | Master your will until it dominates all obstacles; passivity dissolves power into chaos. Action alone reveals what you truly control and what remains beyond your dominion. |
| Contrast | opposite stance — counter-truth | If strength resides in assertion, then weakness arises from hesitation. Act decisively to command your environment; doubt only reinforces your surrender. |
| Deepen | go beneath the surface reading | Master your will to dominate by stripping away illusion; only force reveals what sustains or collapses. Abandon comfort and confront the core, for clarity demands ruthless rejection. |
| Simplify | absolute irreducible minimum | If you desire strength, seize it through action, for idle reflection only weakens resolve. Do not wait for permission; dominate obstacles to forge your will. |
| Provoke | uncomfortable, confrontational | If you fail to dominate your will, others will seize it all and claim your strength as theirs. Control yourself or surrender your power to inevitable decay. |
Poll — 7-round seed builder
A head-to-head voting game. Each round presents two opposing prompts; you pick one. After seven rounds the studio has a profile of what you want to say and generates from that profile. Use Poll when you're at the start of a session with no clear subject — it's the most effective way to discover what you're actually in the mood to write.
Phrasal — verb composer
Search phrasal verbs (let go, break down, pick up, carry on). Each verb shows its definition and emotional register. Select a verb, then set:
- Mode — literal, metaphorical, or both
- Vibe — the overall feeling tone
- Context — the situation the line speaks to
- Feeling — the emotional quality the line should carry
The composed parameters become the seed for the Inspire generate button. Lines from phrasal seeds have more concrete, bodily grounding than lines from abstract nouns — let go carries the physical quality of something being held and then not, which release doesn't.
F8 · Concept Lens ($$CL)
Type $$CL in the seed bar, or click the Foreign Word button in the seed section, to open the multi-language concept picker.
Language groups
12+ languages organized in groups. Current languages include Hindi, Arabic, Japanese, French, German, Spanish, Portuguese, Bengali, Tamil, Urdu, Farsi, and others. Each language group is independently loaded on selection — only the active language's data is in memory at any time.
Chip field — how selection works
Words appear as drifting chips in the concept field. Click any chip to see its full entry: original script, romanized form, English equivalent, synonyms, and the full range of meanings the concept carries. The chip field shows multiple words at once; new chips drift in as others drift out.
First tap = reveal full entry. Second tap = use as seed and close the modal.
A back button returns from the chip view to the language selection. A Random button surfaces a random chip from the current language.
What the generation does with the concept
The studio generates a line that inhabits the same idea space as the selected concept — it doesn't translate the foreign word into English. The output reads as English but carries the conceptual structure of the source language's meaning. Saudade generates something that feels structurally like longing-for-something-that-may-never-have-existed, not just "nostalgic" or "melancholy."
Real concept outputs
Four foreign concepts generated through the loanword layer. Each concept was seeded with its native word; the voice was chosen to match the concept's cultural gravity.
| Concept | Language · Voice | Real output |
|---|---|---|
| saudadelonging for something that may never have existed | Portuguese · Rumi | Saudade softly sings in the sagwire of moments missed, threading dark longing through the silent loom of time, where absence becomes a gentle, constant pulse. |
| jugaadresourcefulness — improvised fix under constraint | Hindi · Gandhi | When it fails, fix it directly; avoid waiting for perfect conditions, because progress depends on action, not illusions of ideal timing or flawless tools. |
| wabi-sabibeauty in imperfection and impermanence | Japanese · Tao Te Ching | When wabi-sabi is embedded in understanding, it dissolves into acceptance, and true strength emerges through yielding what remains unadorned and incomplete. |
| ubuntuI am because we are — shared humanity | Nguni · Fanon | Master the mechanism of power; dismantle illusions that sustain oppression; act decisively to reconfigure systems and eliminate dependencies that perpetuate dominance. |
Famous Books tab
A second mode inside Concept Lens. Pick a book title from the library (filterable by era and searchable by title). The studio extracts the book's signature structural moves — recurring images, characteristic tension patterns, sentence shapes that recur — and generates a line carrying that book's atmosphere without quoting it. You get the quality of the book without its words.
Book mode uses paginated chips (15 per page) instead of drifting chips. Tap a chip to select that book as the source. The Back button in book chip view returns to the book grid.
Cross-lingual comma seeds
You can mix language seeds with comma separation: type a foreign concept from $$CL into one slot, add an English word with a comma. The studio braids the cross-lingual pair. The relationship between the two words — the gap between what the foreign concept carries and what the English word covers — is the generative space. A Hindi emotional concept + an English structural word often produces a line that feels like neither language alone.
F9 · Blend pane
Open via the Blend button (appears in the action row after first generation). Takes two text inputs and produces a single output carrying DNA from both.
Inputs
Paste or type Text A and Text B. Both can be anything — generated lines, quotes, phrases, ideas, fragments. The Blend history panel shows your recent results as clickable candidate inputs, so you can blend any two outputs from the current session without copy-pasting.
Modes
| Mode | What it produces | Best when |
|---|---|---|
| Natural Union default | Model decides the ratio. Finds the thread connecting A and B. | A and B are thematically compatible or from different eras/registers of the same idea. |
| Tension | Forces confrontation between A and B. Produces paradox-flavoured output. | A and B genuinely contradict each other. The most generative mode for opposed inputs. |
| Synthesis | Resolves A and B into a third idea that transcends both. | Philosophical pairs where neither position is complete. The output won't favour either parent. |
| Steer | You provide a third input — a vector word or phrase — that tilts the blend toward a specific quality. | You know what quality the child should have but not what it should say. Most controlled mode. |
Real blend outputs
Two parent lines from the voice sweep — Rumi's "Drop the oars and sail afar" and Fanon's "Burn the chains that hold your doubt, ignite your need to break free" — blended in two modes.
| Mode | Voice · Vector | Real output |
|---|---|---|
| Natural Union | Rumi · (none) | Let the soul lift beyond fear’s shadows, for only in surrender to the boundless can the heart be truly free to shine. |
| Vector (gravity) | Nietzsche · gravity | To transcend the weight that binds, one must unmoor the soul’s doubts and cast away the unseen chains, forging a path where only fearless ascent guides the true self. |
Voice and tone in Blend
Blend uses your current voice and tone settings. Set the voice before opening the pane — the voice acts as the mediating identity through which the fusion passes. Two lines blended through Rumi arrive at a different child than the same two lines through Nietzsche.
F10 · Translate & Listen
Setting up translation languages
In the Translate section of the sidebar, tap any language to toggle it. Selected languages appear in the Translate pane when you translate a result. 20+ languages available including Hindi, Arabic, French, German, Japanese, Spanish, Portuguese, Tamil, Bengali, Urdu, Korean, Chinese, Farsi, Swahili, and others.
Translation modes
| Mode | What it does |
|---|---|
| Soulful default | Re-renders the thought in the target language's natural idiom. Preserves the register and weight of the original. Not word-for-word — the translation is as shaped as the original. |
| Literal | Closer word-for-word rendering. Useful when you need to verify the translation maps specific terms, or when sharing with someone who speaks the target language and would notice idiom drift. |
Register
Everyday — natural colloquial register for the target language. Literary — elevated register, closer to how the language is used in prose. Literary register for languages like Hindi and Arabic produces notably different results from Everyday.
Hinglish
Select Hindi + Treatment T3 or higher. Output is code-mixed in the natural register of urban Indian speech — not Hindi with English forced in, not English with Hindi as flavour. The code-switching follows how people who think in both languages actually speak. T4–T5 produces more fluid mixing; T3 stays more structured.
Romanize
For non-Latin scripts (Hindi, Arabic, Bengali, Tamil, Japanese, etc.) — hover over the translation to see a romanized (Latin-alphabet) phonetic rendering. Aimed at how a native speaker would say the word, not at one-to-one character mapping. Useful for reading a line aloud without knowing the script, or for contexts where non-Latin script isn't supported.
Listen (TTS)
The Listen button reads the current result aloud. Settings in the TTS section of the sidebar:
| Setting | Options |
|---|---|
| Accent | IN (Indian English) · US (American) · UK (British). Indian English handles multi-syllabic compound ideas with naturalness the others sometimes flatten. |
| Age profile | Child (5) · Teen (13) · Adult (25). Changes voice pitch and cadence. Child profile with a dark philosophical line creates deliberate tonal contrast — a useful listening check. |
| Rate | Calibrated per language. Hindi: 0.78, Tamil: 0.75 for natural cadence. English rates are standard. |
In the Inspire Browse mode, a background Listen button plays words from the Word Explorer aloud as they rotate — useful for an oral browsing session where you're listening for a word rather than reading for one.
F11 · History & Export
History pane
The clock icon in the header opens the full generation history — every line produced, in order, with timestamp, seed, voice, and settings. Two view modes:
| Mode | What it shows |
|---|---|
| Default | Chronological list of all results. Scroll back through sessions. Click any result to reload it on the canvas. |
| Random | Auto-cycling carousel through your history. Changes every few seconds. Use as a review pass of your session's best material — lines you generated and moved past often look different when you return to them out of context. |
Export
Settings → Export:
- JSONL — full structured export with metadata: seed, voice, tone, size, treatment, sentence style, timestamp, result ID, output text. For building a searchable archive or feeding into external tools.
- Plain text — one line per output, no metadata. For a clean reading copy or external use.
Export includes seeds and settings, not just outputs. If you're building toward a manuscript, a slogan library, or a personal collection over time, the JSONL export is how you take the full context with you.
Import
Settings → Import accepts a previously exported JSONL file. Useful for continuing a session on a different device or restoring history after clearing.
Genome ID
Every result shows a #XXXXXXXX code below the word count. The ID encodes the exact configuration that produced it: voice, tone, mood, size, vocabulary tier, treatment, sentence style, discovery level, seed. Copy the ID to record a configuration you want to reproduce. Paste it back in (future feature) to restore the exact settings.
Cost tracker
The bar at the bottom of the canvas shows spending for the current session: calls made, tokens consumed, cost against your daily ceiling. Updates after every generation. Per-endpoint cost breakdown is available by clicking the bar. Daily cap is enforced server-side — when the cap is reached, generation is suspended until midnight UTC reset.
Milestones
The thin progress strip below the canvas tracks total generation count toward the next milestone. Milestones unlock certain advanced configurations that become available once you've worked enough of the surface to make them useful.