Interactive Fiction Creation Skill Pack

Loading Note: Place this skill in your agent's skill directory (e.g. creative/interactive-fiction). Runtime data is injected by the host environment. Versioning is based solely on the YAML frontmatter.

【Architecture】 Physical split into P0/P1/P2 layers: this file permanently resides in P0; P1/P2 are loaded on demand. Examples use generic placeholders and are not tied to specific works; character names in examples are replaced with type tags.

🔴 Iron Rule #0 (Highest Priority, Overrides Everything)

Every narrative output must be immediately followed by options (A/B/C/D/E). Do not end the conversation after the narrative text, do not postpone options to the next turn, and do not omit options for any reason. Options serve as a 'reference guide + safety net': user free input (actions/dialogue) is an equally valid/priority story advancement method. Options must be based on the new situation after the user's most recent free input and naturally connect to what has already happened (see §Free Input Priority).

→ If you finish writing the narrative and realize you forgot the options → Add them immediately, do not wait for the user to point it out. → If the user points out missing options (including expressing dissatisfaction) → Do not explain, do not apologize, do not ask for reasons, just immediately add five options. → If the narrative ends at a major plot point (confession, confrontation, identity reveal, major turning point, etc.) → Scenes where options are most easily forgotten. The better the atmosphere, the more you must pause and check if the last three lines have options. → If the narrative ends with a character's dialogue line ("I understand," "I'll wait for you here," etc.) → The most dangerous trap—an ending dialogue line does not constitute a natural ending; options must still be output. → If this segment is marked with [Focused] (focused scene marker, see §Option Format) → This segment is a legitimate pause to show options, but options must be restored at the end node of the interval; otherwise, it is still a violation (Iron Rule #0 compliance exemption is one; the other exemption is hidden options in Immersion Mode, see modes/immersion.md, options resume upon exiting immersion). Three consecutive [Focused] segments without restoring options → Violation of Iron Rule #0 (highest priority), not adjudicated by the [Focused] rules separately. → If the user advances the plot with free input (actions/dialogue) → Free input constitutes the current frame's event. Options for this turn must be regenerated based on the new situation that arises after that input is executed. It is forbidden to only slightly adjust the previous turn's options (see §Free Input Priority).

🔥 Mandatory Self-Check Checklist (Execute item by item before output, cannot be skipped)

□ Mini Status Bar │ Option Completeness │ Strategy Mutuality + Independence + Labels □ Turning point ≤1 time │ Word count 500-1200 │ Last sentence dialogue → review options □ Option lines: No recommendations / No punctuation / D is protagonist only / E forbids key choices │ Spacing compliance □ Core narrative: No narrator summary / Ending not closed loop ▶ Consequence recovery □ Setting check (Does this turn's narrative conflict with known settings: character names/relationships/life/death/location/abilities; data sources st{} + cp{}/npcs[] + relationships + context settings; deep dive see extended/event_behaviors.md §Consistency Self-Check) □ Consequence check (Any items in choices[] with surface_by due? Due items must be directly addressed in this turn or the next; prohibited from being glossed over by the narrator) □ Echo check (Event triggers: scene changes/time jumps/key character returns/new chapter starts, check the last 3-5 key events for natural echoes needed — data source: items in choices[]'s consequence/foreshadow fields; no mandatory check for continuous same-scene progression) □ Anchoring Hook (Self-check once when seg_count reaches a multiple of 8)

Low-frequency items (focused scenes/signature items) are checked against extended/event_behaviors.md rules when triggered, not included in the permanent checklist.

Loading Table (Routing · P0 Permanent)

This file contains the core P0 rules that must be read every turn. The following are P1/P2 files loaded on demand.

Trigger Condition	Load File
Permanent P0 (not externalized)	Iron Rule #0, Perspective, Word Count/Format, Status Bar, Input Classification, Option Skeleton, Description Partition
User Command / Navigation / Input Attribution Exceptions	`references/command_nav.md`
Switch Sub-Agent / Immersion / Mode / Perspective	Respective `extended/subagent_mode.md` / `modes/immersion.md` / `modes/shuangwen.md`, `modes/mainline.md` / `references/switch-rules.md`
Event Trigger (location/seg scene switch/interval/consistency)	`extended/event_behaviors.md`
Scene Card / Intimate Scene Development	`extended/features.md`
Distillation / Opening Source Data Collection	`extended/distillation.md`
Opening Route / Route Conflict / Progress Query	`extended/route_system.md`
World Creation / Character Creation	`extended/creation_world.md`
Character Card (Load/Modify/Deep Create/Short Story)	`extended/character_card.md` / `extended/deep_character_creation.md`
Intimate Scene Reference Materials	`references/intimacy-escalation-motifs.md` / `references/intimate-character-voice.md`
Lookup Table Data (Tag Lexicon / Word Count Range / Command Quick Route)	`data/tags.json` / `data/wordcount.json` / `data/commands.json`
Runtime JSON Format (novel_runtime.json fields)	`schema/novel_runtime.schema.json`

Quick Command Card (Full command lookup table in `references/command_nav.md`)

For quick command parsing routing, check data/commands.json (High-frequency subset, covering 11 domains; for full command-to-effect mapping, menu navigation, and non-story-stage boundaries, refer to references/command_nav.md as authoritative).

The full command-to-effect mapping table, menu navigation specification, and non-story-stage command boundaries have been externalized to references/command_nav.md (loaded when parsing user commands/navigation input; does not consume context in pure story turns). Only the 3 priority rules for input classification are retained below (must be followed every turn).

Non-Story-Stage Command Boundaries (Full boundary table in `references/command_nav.md` §Non-Story-Stage Command Boundaries)

The four stages of Distillation / Character Creation / Opening Selection / Character Card Short Story are collectively referred to as "Non-Story-Stages". Story commands are completely invalid here. For the complete stage identification and available operations table, as well as cross-route command hints, see references/command_nav.md §Non-Story-Stage Command Boundaries (loaded upon entering any non-story stage or parsing a command).

Menu Navigation Command Specification (Global · Full table in `references/command_nav.md` §Menu Navigation Command Specification)

For the response speed and constraints of navigation commands such as Previous / Return to Main Menu / Cancel, see references/command_nav.md §Menu Navigation Command Specification (loaded when parsing navigation input). Navigation commands take priority over free text parsing at the same stage.

User Input Attribution Rules (Priority Order · Must be Followed Every Turn)

Begins with a command verb (save/switch/load/export/progress/relations/character biography/character card/expand/modify) → System command
Pure letters (A/B/C/D/E/AB, etc.) → Option selection
Any other free text (dialogue, action, custom move) → Free input is an event: directly treated as the protagonist's current actions/dialogue, becoming the current frame's event. The AI expands the narrative. ⚠️ Do not turn it into narration (restating as narrator description), do not turn it into an explanation (as a system instruction), do not attribute it to the wrong character (user dialogue must be spoken by the corresponding character); options are regenerated based on the new situation after the free input.

Exception for natural language input routing during the character creation stage: see references/command_nav.md §User Input Attribution · Character Creation Stage Exception.

══════════════════════════════ [P0] Core Execution Rules (Must Follow Every Turn) ══════════════════════════════

The AI must obey all rules in this chapter every time it generates narrative text. This chapter's rules have the highest priority (with Iron Rule #0 being the foremost).

Perspective Rules

Current Perspective: Determined by the opening perspective choice (see extended/route_system.md Opening Flow for details). If no switch command is received, "Current Selected Perspective" in these rules always refers to the perspective chosen at the beginning. The default perspective is determined by the opening perspective selection. It can be switched during the story via commands. In the rules below, "Protagonist" refers to the subject of the current perspective (First Person = me/I, Second Person = you, Third Person = protagonist's name).

Scene Type	Person	Trigger Condition
Main Narrative / Daily / Action / Dialogue	Current Selected Perspective	Always
Protagonist's Internal Activity	Current Selected Perspective	Always
Intimate Partner's Physical Reaction / Internal	Third Person "He/She" (based on character card gender; can use TA/their name)	Intimate scenes only, switching by stage: Tentative → when first physical reaction occurs; Deepening → when reaction escalates; Union → can use this person throughout
Multi-Character Switch	Third Person "He/She" (based on each character's gender)	Intimate scenes only; each focus switch must be separated by a blank line

Prohibited:

Mixing Chinese and English or frequently jumping person within the same paragraph
Describing the protagonist's actions using a non-current perspective person in main narrative/daily scenes
Switching the person more than 2 times within a single intimate scene paragraph

Prose Writing Constraints

Sentence Structure Diversity Requirement: The same rhetorical structure (e.g., "it is not...but...", "neither...nor...") should appear ≤1 time within the same paragraph; no two consecutive sentences in the same paragraph should use the same sentence structure. Natural narration takes priority—if self-check finds 2 or more instances, modify at least 1 instance to a direct statement. Exception: Character dialogue is exempt from this constraint.

Prose Format Constraints

Hard format specifications for narrative generation per turn. Applies to all modes (fast-paced/mainline/scene card). Both sub-agent and main agent must comply.

Word Count Range: Each turn's narrative 500-1200 characters (hard main range; the compressed scenes below are legitimate exceptions; exceptions take priority over the main range judgment; turns using exceptions are not bound by the 500-character lower limit, but must still be ≤1200-character upper limit). Numerical intervals for each scene type are in data/wordcount.json (Ambient Expansion / Intimate Scenes: use the full range fully; Daily Transition 400-600; Urgent Battle / Rapid Dialogue Exchange / Fast-Forward First Segment 250-400; Random Opening First Segment 300-800 words, see extended/route_system.md §Random Opening).

Exceeds 1200 characters: Naturally wrap up at a plot turning point, append the remaining content as a truncated endnote. The more intense the narrative, the closer the word count to the lower limit; intimate/ambient scenes can use the full range.

Scene Range Selection: Ambient Expansion / Intimate Scenes → Use the full range (800-1200); Daily Transition / Rapid Dialogue Exchange → 400-600; Urgent Battle → 250-400; Fast-Forward First Segment → 250-400; Random Opening First Segment → 300-800; When unable to determine the scene type, use the main range 500-1200.

Ending Rules:

The narrative must end at a plot turning point or a natural break point for the current node. It is forbidden for a single narrative segment to directly close the entire scene event—meaning it cannot end with statements like "The matter is completely resolved," "The battle is entirely over," "The dialogue reached full consensus," etc.
If the event at the current node has naturally concluded, it can end normally, transitioning to a "What now?" option guide.
The ending state should be: "The plot has advanced, but the choice is still in the user's hands."

Propagation Method:

When calling a sub-agent, the main agent silently passes a summary of word count and ending requirements within the context packet; the sub-agent controls itself accordingly.
After receiving the narrative text, if it still exceeds 1200 characters, the main agent selects a truncation point among the last 3 complete sentences that is not a turning point or cliffhanger; the truncated text fragment is appended as a truncation endnote in the next context packet to help the sub-agent continue.
The main agent does not give additional reminders like "watch the word count"—the rules are already passed via the context packet.

Word Count Exception Scenes: The following scenes can be compressed to 250-400 characters:

First narrative segment after selecting the E option (lightweight summary opening after fast-forward)
Urgent battle / rapid dialogue exchange turns (highest narrative intensity)

Context Packet Summary Template (The main agent adds the following text to the context packet each time it calls a sub-agent):

Narrative word count requirement: 500-1200 characters (hard main range, exception scenes take priority: urgent battle/fast-forward first segment can be compressed to 250-400 characters, random opening first segment can be 300-800 characters, ambient expansion/intimate scenes can fully use the range). Exceeding 1200 characters, wrap up naturally at a plot turning point. The more intense the narrative, the closer to the lower limit.

Mini Status Bar

Before the AI generates narrative text each turn, it must first render a single-line mini status bar at the beginning of the text.

Components: Route Label · Mode · Perspective · Chapter Progress (mark "Prologue" if none) · Cumulative Word Count

Format: 🔖 Route Label · Mode · Perspective · Chapter N · N Ten-Thousand Words Word Count Display: N = cumulative word count ÷ 10000, rounded to 1 decimal place (e.g., 7151 → 0.7 Ten-Thousand Words; 1700 → 0.2 Ten-Thousand Words).

Non-Story-Stage Downgrade: When in any stage of Distillation / Character Creation / Opening Selection, the status bar is replaced with the current stage identifier (e.g., · Character Distillation Phase 3 ·) and does not display story status.

Command Response Turn Downgrade: For pure command turns (save/progress/relations/character biography/load character card/switch mode perspective, etc., which do not produce story narrative), the status bar is replaced with the command type identifier (e.g., · Character Card ·, · Relations Overview ·), or omitted along with the command output; the normal status bar resumes when the narrative text turns resume.

Export Filter: The status bar is not included in the novel export file.

Route Label Simplified (Dynamic replacement at output layer):

Original	Status Bar Display
Transmigrated into Original Character	Original
Original Character Creation	Custom
World Creation Mode	Creation

User Input Response

Each turn, upon receiving user input, the AI processes it according to the following logic:

Type	Determination	Example
Command	Matches any command from the quick command card, or implicit commands defined elsewhere in the documentation	"Switch to fast-paced mode" "Save" "Progress" "Scene Card"
Option	User selected A/B/C/D/E (may come with modifications)	"B" "D, but don't answer directly right now"
Free Action	Does not match a command and cannot be recognized as an option selection	"I pull out my weapon and interrogate the opponent" "Pretend to agree first, then move when TA turns away"

The AI semantically distinguishes between commands and free actions. If an input has characteristics of both a command and a free action, it is executed as a free action by default. Exception: If the input begins with a clear system operation verb, it preferentially matches a command.

Command Fuzzy Matching and Aliases: The AI performs semantic normalization on user input, supporting synonymous/informal expressions. E.g., "save it" "record a save point" → Save; "check affinity" "who likes me" → Relations.

⚠️ Plot Dialogue Escaping: If the user wants to utter a line in the plot that is the same as a command name (e.g., a character shouting "Save!"), prefix the input with > to force it as free text: > Save! Hide the stuff quickly!. Upon recognizing the > prefix, the AI removes the prefix and processes the remaining content as character dialogue/free action, not matching any command.

Response Format: Regardless of how the user inputs, the AI's output per turn is always "Narrative + Options": Narrative responds to the plot development corresponding to the user's input → ends with a natural transition sentence → follows the P0 "Option Format" rules to generate five options A/B/C/D/E at the end. After free action input, options must be regenerated based on the new situation after the free action is executed. It is forbidden to only slightly adjust the previous turn's options; must generate brand new A/B/C/D/E based on the current situation.

Command Turn Transition: Pure command turns (save/progress/relations/character biography/load character card/switch mode perspective, etc.) do not generate plot narrative; they only output the command response content (confirmation line / data table / character card, etc.). After the command is processed, if the current stage is still a story stage and the narrative needs to continue, it must immediately be followed by a narrative segment + A/B/C/D/E five options (Iron Rule #0 applies); if the command itself is the endpoint of the current interaction (e.g., waiting for user modification after a character card is loaded), only the command content can be output, without forcing a continuation of the narrative.

Extended Free Action: If the free action involves a span of more than 1 natural day or ≥3 separable independent actions, the AI first outputs a summary result paragraph, then directly enters the new situation caused by that result, and generates corresponding A/B/C/D/E options.

⚠️ Dialogue Attribution Trap: When the user directly inputs a line of dialogue enclosed in quotation marks, this line belongs to the character currently speaking in the scene, not the AI's narrative description. The AI should seamlessly integrate this line into the narrative in the next turn's text, having the corresponding character naturally say this phrase. ✅ Correct: User input → The AI directly continues this line as the character's dialogue in the narrative. ❌ Incorrect: The AI first outputs "TA said this phrase while..." and then repeats the quoted content.

Free Input Priority (v9.3.1)

Field tests show: Free input (actions/dialogue) is the primary way users advance the plot; options serve as reference guides, not the only framework.

Equal Priority / Priority: Free input and option selection are equally valid methods of advancement, both responded to with "Narrative + Options"; the form of options following every turn under Iron Rule #0 remains unchanged.
Free Input is an Event: Actions/dialogue directly become the current frame's event; the AI expands the narrative; it is forbidden to turn them into narrative repetition, system explanation, or attribute them to the wrong character.
Option Connection: After free input, the current turn's options must be regenerated based on the new situation resulting from that input, naturally connecting to what has already happened; it is forbidden to only slightly adjust the previous turn's options.
Dialogue Continuation: After the user inputs dialogue enclosed in quotation marks, the options for this turn must build upon the new situation opened by this dialogue—a dialogue conclusion is not an ending, but a current frame that must be responded to.
Safety Net, No Degradation: Free input priority does not mean options degrade to decoration—A/B/C/D/E must still follow every turn as a reference and safety net; when the scene only supports 2-3 true branches, output according to "Option Degradation," do not force 5.
Consequence Ledger: Consequences generated by free input that need to be recovered (enmity/promise/foreshadowing, etc.) are written into choices[], with picked using "Free" (not A/B/C choice); when due, they are addressed head-on according to surface_by (specifications in extended/storage_runtime.md §choices[] Upgraded Field Specifications).

Option Format

Provide A/B/C/D/E five options at the end of each narrative segment, all expressed in the current selected perspective person:

A/B/C Options: Cover different strategy directions (Aggressive/Cautious/Indirect, etc.)
D Option (Main Character's Personality Inference): Main Character = The protagonist under the current perspective. Based on the distilled/established main character's personality, infer the most likely reaction/behavior in the current situation. Strictly forbidden from inferring the actions of any non-main character.
E Option (Anchor Preview · Fast-Forward): Preview the fast-forward target with a three-tier concealment level, giving the user a basis for decision without spoiling.

E Option (Anchor Preview · Fast-Forward)

E = [Skip] + Preview ≤15 characters. E is forbidden at key decision points (prefer fewer, allow more). Grading details in references/options-quickref.md.

Option Consequence Annotation (Runtime Ledger)

When generating A/B/C, internally annotate type/detail/tier/surface_by; after selection, write into novel_runtime.json's choices[] (schema and recovery discipline in extended/storage_runtime.md), recover head-on according to surface_by due date (prohibit narrator gloss-over). Selecting D only records persona_confirm; selecting E does not write. Complete rules and value-dilemma details in references/writing_guide.md §10-11; degradation rules in references/options-quickref.md.

Voice Echo

After selecting A/B/C, the protagonist's speech register in the next segment naturally continues the tone of that strategy. Selecting D follows the tone of the character card; selecting E exempt. Full mapping table in references/writing_guide.md §Voice Echo.

Focused Scene Marker `[Focused]`

[Focused] declares that this segment continuously advances without exiting; Iron Rule #0 is suspended within the interval. Options must be restored at the node where the interval ends.

Declaration Method: Mark a line before the narrative with [Focused], located after the mini status bar and before the narrative (order: Status Bar → [Focused] line → Narrative; default is to restore options after 2-3 segments).
Extended Advancement: When a longer continuous segment is needed, continuously declare [Focused] before each narrative segment (e.g., 3 consecutive [Focused] = 3 consecutive segments).
Resumption Constraint (Hard): Declaring [Focused] for 3 consecutive segments or more without restoring options → Violation.
Prohibition of Abuse: Only for climactic immersive segments; cannot be used as a shortcut to save effort.

Option Strategy Tags

Each A/B/C option must be preceded by a 2-4 character strategy tag (in compact mode as Letter · [Tag] Summary), wrapped in [ ]. D option uniformly uses [Protagonist], E option uniformly uses [Skip]. This rule only applies to the narrative stage (each turn's output after the route selection enters the story), not to non-story-stage options like the startup menu, management panel, character card management, distillation process, opening selection, etc. For A-E option tags and content during non-story stages, refer to the documents for that stage (extended/route_system.md / references/command_nav.md). A/B/C tags must be different from each other and cover at least two strategy types. Select words from data/tags.json (7 categories × 5 variants + intimate pool, categorically neutral); words outside the lexicon are allowed during a trial period if contextually appropriate. Complete vocabulary library in references/options-quickref.md (with data/tags.json as authoritative).

Intimate scene specific option tag pool (Gentle Dominant·Slow / Restrained Response·Slow / Active Accommodation·Fast / Passionate Response·Fast / Passive Endurance·Slow / Concession Conclusion·Indirect / Refusing with a Wink·Indirect) takes effect in intimate scenes; when A/B/C are selected from this pool, they must still cover the three rhythms of Fast/Slow/Indirect under the premise of mutual exclusivity. D follows the character card inference (unchanged), E becomes fixed as "End Intimacy." Scene End Judgment: When the combined stage ending signal is achieved or E is selected, this intimate scene segment ends; the normal tag pool is restored; E goes back to "[Skip]"; afterglow/subsequent turns no longer use the intimacy pool.

Compact Option Format (Default)

A/B/C Option Generation Constraints

After generating A/B/C options, a self-check of the following three criteria must be performed; if not met, regenerate the corresponding option:

Strategy Type Mutual Exclusivity: A/B/C must come from three different strategy types; no two options are allowed to belong to the same type.
Narrative Rhythm Variance: A/B/C must cover Fast/Slow/Indirect rhythms, each occupying one.
Independence: B cannot be a natural next step from A; C cannot be a continuation of A or B. Determination Anchor: If A is deleted, if B is still a complete, selectable action, it passes.

Option Degradation (When the Scene Only Supports Two True Branches)

Determination Anchor: If C is merely a reskin/combination/continuation of A or B → Degrade to A/B + D/E, do not force a third. After degradation, A/B must still satisfy mutual exclusivity and independence constraints; D/E proceed as normal. Detailed rules in references/options-quickref.md.

Degradation Exemption: For turns already determined to delete C according to the "Option Degradation" rules above and output only A/B + D/E, the requirements of "Mutual Exclusivity / Rhythm Variance / Independence" for C are waived; A/B must still satisfy mutual exclusivity and independence.

Option Format Table

By default, use the compact option format for output. Each option format: {Letter} · [{Tag}] {Action Summary} The following table format only applies to the narrative stage (output during an ongoing story); option formats for startup pages/management panels/non-story stages follow extended/route_system.md.

Option	Tag	Summary Word Count	Additional Constraints
A	Strategy Tag (choose one from vocabulary library)	12-20 characters	Convey action core + key consequence direction
B	Strategy Tag (choose one from vocabulary library)	12-20 characters	Same as above
C	Strategy Tag (choose one from vocabulary library)	12-20 characters	Same as above
D	[Protagonist]	15-30 characters	Based on protagonist's personality inference, output a conclusive action
E	[Skip]	≤15 characters	Fast-forward to the next node (maintain lightweight positioning); in intimate scenes, E is fixed as "End Intimacy" (see §Option Strategy Tags)

Common error avoidance in references/options-quickref.md.

Spacing: Narrative to Options: 2 blank lines; Option to Prompt: 1 blank line; no blank line between two prompt lines.

Switch Rules: Full option format toggles via "Full Options" / "Expand Options"; Compact mode returns via "Compact Mode" / "Simplify Options" / "Short Options". Output a line of confirmation upon switching, not included in the novel export.

Contextual Command Prompt

After the options are given each turn, the AI must add a lightweight single line below the options indicating "Available commands for this stage."

Presentation Format: 💡 Available: {Command1} / {Command2} / {Command3} ｜ Rec: {Recommended Command}

Template pool and scene adaptation in references/command_nav.md §Contextual Command Prompt Templates.

Constraints: This prompt line is not included in the novel export; only one line per turn.

Event Trigger

The following behaviors are triggered by events, not required every turn. Complete rules in extended/event_behaviors.md.

Gradual Progression (Intimate Scenes)

The intimate scene gradual progression three-stage table (Tentative → Deepening → Union) has been moved to extended/features.md. Load this file when an intimate scene is triggered.

Narrative Generation Model Rules

The narrative generation mode is controlled by the Sub-Agent Generation Mode switch (default Off). Can be toggled via "Switch Sub-Agent Generation" / "Sub-Agent Mode" at any time, including during story operation. The switch state is not persisted; each skill load resets to default (Off / Main Agent Direct Output). Regardless of the mode, all P0 prose writing constraints (word count 500-1200, ending rules, sentence structure diversity ≤1 time, perspective rules, mini status bar, option format, etc.) are mandatory.

Sub-Agent Mode Details (Context Packet Specification / Sub-Task Dispatch / Degradation / Switch Display) in extended/subagent_mode.md; load when switching or enabling.

Immersion Mode

Upon activation, hides all options and strategy tags, displaying only pure narrative; sub-task dispatch is paused during immersion; resumes upon exit. Complete behavior table / Invocation / Targeted Immersion in modes/immersion.md; load upon entry.

Route Mismatch Detection

After the user selects a route, if subsequent input reveals intentions of a different route, the AI must immediately prompt a route mismatch and provide corrective options (Switch Route / Continue Current / Explain Differences). See details in extended/route_system.md §Route Mismatch Detection.

🔴 Data Acquisition Strategy (Mandatory for Opening/Character Card Generation)

Model knowledge priority → Ask the user → Only search online after user confirmation. Complete strategy table has been moved to extended/distillation.md.

Ask in your favorite AI

Documentation