Improve how instructions are provided to high-level fns

src/marvin/fns/cast.py

@@ -22,9 +22,7 @@
 - When providing a string response, do not return JSON or a quoted string
     unless they provided instructions requiring it. If you do return JSON, it
     must be valid and parseable including double quotes.
-- When converting to bool, treat "truthy" values as true
-
-"""
+- When converting to bool, treat "truthy" values as true"""
 
 
 async def cast_async(
@@ -69,14 +67,18 @@ async def cast_async(
         True
 
     """
+    if target is str and instructions is None:
+        raise ValueError("Instructions are required when casting to str")
+
     task_context = context or {}
     task_context["Data to transform"] = data
+    prompt = PROMPT
     if instructions:
-        task_context["Additional instructions"] = instructions
+        prompt += f"\n\nYou must follow these instructions for your transformation:\n{instructions}"
 
     task = marvin.Task[target](
         name="Cast Task",
-        instructions=PROMPT,
+        instructions=prompt,
         context=task_context,
         result_type=target,
         agents=[agent] if agent else None,

src/marvin/fns/classify.py

@@ -15,8 +15,7 @@
 as possible when labeling text. You use inference or deduction whenever
 necessary to understand missing or omitted data. Classify the provided `data`,
 text, or information as one of the provided labels. For boolean labels,
-consider "truthy" or affirmative inputs to be "true".
-"""
+consider "truthy" or affirmative inputs to be "true"."""
 
 
 @overload
@@ -106,9 +105,11 @@ async def classify_async(
 
     """
     task_context = context or {}
-    task_context["Data to classify"] = data
+    task_context.update({"Data to classify": data})
+
+    prompt = PROMPT
     if instructions:
-        task_context["Additional instructions"] = instructions
+        prompt += f"\n\nYou must follow these instructions for your classification:\n{instructions}"
 
     # Convert Enum class to sequence of values if needed
     if labels is bool or issubclass_safe(labels, enum.Enum):
@@ -121,7 +122,7 @@ async def classify_async(
 
     task = marvin.Task[result_type](
         name="Classification Task",
-        instructions=PROMPT,
+        instructions=prompt,
         context=task_context,
         result_type=result_type,
         agents=[agent] if agent else None,

src/marvin/fns/extract.py

@@ -18,8 +18,7 @@
 - When providing integers, do not write out any decimals at all
 - Use deduction where appropriate e.g. "3 dollars fifty cents" is a single
     value [3.5] not two values [3, 50] unless the user specifically asks for
-    each part.
-"""
+    each part."""
 
 
 async def extract_async(
@@ -60,12 +59,13 @@ async def extract_async(
 
     task_context = context or {}
     task_context["Data to extract"] = data
+    prompt = PROMPT
     if instructions:
-        task_context["Additional instructions"] = instructions
+        prompt += f"\n\nYou must follow these instructions for your extraction:\n{instructions}"
 
     task = marvin.Task[list[target]](
         name="Extraction Task",
-        instructions=PROMPT,
+        instructions=prompt,
         context=task_context,
         result_type=list[target],
         agents=[agent] if agent else None,

src/marvin/fns/generate.py

@@ -24,8 +24,7 @@
 If the user provides additional instructions or a description, assume they are
 looking for examples that satisfy the description. Do not provide more
 information than the user requests. For example, if they ask for various
-technologies, give their names but do not explain what each technology is.
-"""
+technologies, give their names but do not explain what each technology is."""
 
 
 async def generate_async(
@@ -62,12 +61,13 @@ async def generate_async(
 
     task_context = context or {}
     task_context["Number to generate"] = n
+    prompt = PROMPT
     if instructions:
-        task_context["Additional instructions"] = instructions
+        prompt += f"\n\nYou must follow these instructions for your generation:\n{instructions}"
 
     task = marvin.Task[list[target]](
         name="Generation Task",
-        instructions=PROMPT,
+        instructions=prompt,
         context=task_context,
         result_type=conlist(target, min_length=n, max_length=n),
         agents=[agent] if agent else None,

src/marvin/fns/summarize.py

@@ -17,8 +17,7 @@
 - Focus on the most important information
 - Use clear, concise language
 - Adapt tone and style to match the content type
-- Ensure the summary stands alone as a coherent piece of text
-"""
+- Ensure the summary stands alone as a coherent piece of text"""
 
 
 async def summarize_async(
@@ -59,12 +58,15 @@ async def summarize_async(
     """
     task_context = context or {}
     task_context["Data to summarize"] = data
+    prompt = PROMPT
     if instructions:
-        task_context["Additional instructions"] = instructions
+        prompt += (
+            f"\n\nYou must follow these instructions for your summary:\n{instructions}"
+        )
 
     task = marvin.Task[str](
         name="Summarize Task",
-        instructions=PROMPT,
+        instructions=prompt,
         context=task_context,
         result_type=str,
         agents=[agent] if agent else None,

src/marvin/templates/task.jinja

@@ -8,7 +8,9 @@
     </instructions>
     {% if task.context %}
     <context>
-        {{ task.context }}
+        {% for key, value in task.context.items() %}
+        <{{ key }}>{{ value }}</{{ key }}>
+        {% endfor %}
     </context>
     {% endif %}
 

tests/ai/fns/test_cast.py

@@ -93,7 +93,7 @@ def test_cast_text_with_subtle_instructions(self):
         result = marvin.cast(
             "My name is marvin",
             str,
-            instructions="Rewrite with uppercase names e.g. JOHN, but leave all other text unchanged",
+            instructions="Change only names to uppercase (e.g. JOHN), but leave all other text unchanged",
         )
         assert result == "My name is MARVIN"