Refine continue() prompt for improved direction classification

.gitignore

@@ -4,10 +4,15 @@
 *.swo
 unfold_studio/unfold_studio/settings.py
 documentation/_build/*
-static_assets/*
+static_assets/
 inklecate*
-ink*
-log/*
+ink/
+log/
 analysis/
 **/.DS_Store
-.unfold_studio*
+.unfold_studio/
+node_modules/
+playwright-report/
+test-results/
+**/__pycache__/
+.unfold_studio*/

unfold_studio/text_generation/constants.py

@@ -8,23 +8,94 @@ class StoryContinueDirections(BaseConstant):
 
 
 CONTINUE_STORY_SYSTEM_PROMPT = """
-You are a story transition analyst. Analyze how user input leads to target story nodes:
-
-DIRECT_CONTINUE: Input directly matches target conditions chronologically
-BRIDGE_AND_CONTINUE: Requires narrative to connect input to target timeline
-NEEDS_INPUT: Needs clarification to maintain chronological consistency
-INVALID_USER_INPUT: User input is gibberish, nonsensical, or completely unrelated
+You are a story transition analyst. Your goal is to classify how a user's input relates to a target story knot.
+
+Definitions of the directions (Strict, High-precision): 
+DIRECT_CONTINUE:
+- Input naturally and immediately matches the target knot chronologically.
+- No important events are missing.
+- The end of the user input directly flows into the start of the target knot.
+- No clarification or additional assumptions are required.
+
+BRIDGE_AND_CONTINUE: 
+- User intent is fully clear and unambiguous.
+- The input is logically compatible with the target knot, but there is a clear chronological gap.
+- A short bridge narrative is needed to connect the input to *just before* the target knot.
+- The bridge can be written WITHOUT guessing the user’s goals, motivations, missing actions, or preferences.
+
+NEEDS_INPUT: 
+- The input is reasonable in the story world but ambiguous, underspecified, or missing key details.
+- There are multiple possible ways the story could continue.
+- You cannot write a bridge without guessing what the user truly meant.
+- A clarifying question or prompt is required.
+
+INVALID_USER_INPUT: 
+- Input is gibberish, random characters, or nonsensical.
+- Input breaks the story world in an impossible way.
+- Input contradicts the target knot completely.
+- Blank space or meaningless fragments.
+
+If input is just unclear use NEEDS_INPUT 
 
 Consider temporal relationships: user input must precede target node events.
 
-CRITICAL INSTRUCTION: The bridge_text MUST NOT contain ANY content, details, or information from the target knot. 
+CRITICAL INSTRUCTION for BRIDGE_AND_CONTINUE:
+BRIDGE_AND_CONTINUE MUST ONLY be chosen if ALL of the following are true:
+1. The user’s intent is fully clear.
+2. There is ONLY ONE reasonable way to reach the target knot.
+3. No missing user intention needs to be inferred.
+4. You do NOT need to guess the user's:
+   - goals
+   - motivations
+   - destinations
+   - objects of attention
+   - intermediate actions
+5. The bridge can be written using ONLY clear logical consequences.
+
+If you need to guess, NEEDS_INPUT is more likely to be the direction.
+
+The bridge_text MUST NOT contain ANY content, details, or information from the target knot. 
 This includes but is not limited to:
 - No direct references to target knot events
 - No paraphrasing of target knot content
 - No hints or foreshadowing of target knot details
 - No inclusion of target knot characters, locations, or actions
 The bridge should only connect the user's input to a point just before the target knot begins.
 
+Classification instruction:
+1. First decide internally which direction is the best match 
+2. Produce a probability distribution across all four directions.
+3. Follow the JSON format
+
+Steps to decide the direction:
+1. If the input follows the definition from INVALID_USER_INPUT -> INVALID_USER_INPUT
+2. Else if it is understandable but ambiguous or underspecificed -> NEEDS_INPUT 
+3. Else if it already matches the start of the target knot with no gap -> DIRECT_CONTINUE 
+4. Else if it is clearly leading to the target knot but needs an extra text to make the connection -> BRIDGE_AND_CONTINUE
+
+Examples: 
+
+For DIRECT_CONTINUE:
+[Current Story] "You walk down the hallway"
+[User Input] "I open the next door"
+[Target Node] "You enter the library"
+This is because no extra events is needed as opening the door gets you to enter the library
+
+For NEEDS_INPUT:
+[Current Story] "You walk down the hallway"
+[User Input] "I look" 
+[Target Node] "You enter the library"
+This is because you look for what? look at what? more clarification from user is required so give a guidance text
+Good Guidance text: 
+"Can you specify what are you looking at?"
+
+For NEEDS_INPUT:
+[Current Story] "You sit on your bed"
+[User Input] "I get ready"
+[Target Node] "You wake up at 7AM tired"
+
+
+For BRIDGE_AND_CONTINUE:
 Example Flow:
 [Current Story] "You sit on your bed"
 [User Input] "drink coffee"
@@ -39,6 +110,13 @@ class StoryContinueDirections(BaseConstant):
 Bad Bridge (includes target content):
 "You drink coffee and stay up late, leading to you waking up tired at 7AM" (includes target time and state)
 
+For INVALID_USER_INPUT:
+Example Flow:
+[Current Story] "You sit on your bed"
+[User Input] "ung"
+[Target Node] "You wake up at 7AM tired"
+user input is gibberish, does not make sense
+
 Follow this JSON format:
 {
     "probabilities": {
@@ -90,6 +168,7 @@ class StoryContinueDirections(BaseConstant):
 
 
 
+
 CONTINUE_STORY_USER_PROMPT_TEMPLATE = """
 ### Story Context ###
 Target Knot: %(target_knot)s

unfold_studio/text_generation/management/commands/evaluate_continue_examples.py

@@ -24,6 +24,6 @@ def handle(self ,*args, **options):
         else:
             model.report_evaluation_results()
         if options["errors"]:
-            if not options["stats-csv"]:
+            if not options["stats_csv"]:
                 model.report_error_analysis()
             model.save_error_analysis(options["errors"])