prompt-builder-py

Fluent prompt builder for LLM APIs — templates, variable substitution, token counting, and chat history management in one zero-dependency Python library.

Installation

pip install prompt-builder-py

Why prompt-builder-py?

• Fluent API — chain .system(), .user(), .assistant(), and .var() calls naturally.
• Variable substitution — {name} placeholders resolved at build time, not at declaration.
• Handlebars templates — {{variable}}, {{#if}}, {{#each}}, partials, and a global registry.
• Chat history — add messages, auto-trim to token/count limits, export to JSON/Markdown.
• Provider formatters — emit the exact payload shape for OpenAI, Anthropic, Cohere, Mistral, or plain text.
• Token counting — dependency-free approximation using a hybrid char/word heuristic.
• Zero dependencies — pure Python 3.9+, stdlib only.

---

Quick Start

from prompt_builder_py import PromptBuilder, OpenAIFormatter

messages = (
    PromptBuilder()
    .system("You are an expert in {domain}.")
    .user("Explain {concept} with a short example.")
    .var("domain", "Python")
    .var("concept", "decorators")
    .build()
)

payload = OpenAIFormatter(model="gpt-4o").format(messages, max_tokens=512)
# → {"model": "gpt-4o", "messages": [...], "max_tokens": 512}

---

PromptBuilder

Building messages

from prompt_builder_py import PromptBuilder

builder = (
    PromptBuilder()
    .system("You are a {role}.")
    .user("What is {topic}?")
    .assistant("Here is what I know about {topic}…")
    .user("Can you elaborate?")
)

Variable binding

# Bind one at a time
builder.var("role", "historian").var("topic", "the Roman Empire")

# Or bind many at once
builder.vars(role="historian", topic="the Roman Empire")
builder.vars({"role": "historian", "topic": "the Roman Empire"})

Strict mode

# Raise VariableNotFoundError for any unbound {variable}
builder.strict()

from prompt_builder_py import VariableNotFoundError
try:
    builder.build()
except VariableNotFoundError as e:
    print(e.variable_name)

Forking a builder

base = (
    PromptBuilder()
    .system("Answer in {language}.")
    .user("{question}")
    .strict()
)

en = base.copy().vars(language="English", question="What is entropy?")
fr = base.copy().vars(language="French",  question="What is entropy?")

print(en.build())
print(fr.build())

Output methods

Method	Returns
.build()	list[dict] — all messages as {role, content} dicts
.build_chat()	list[dict] — non-system messages only
.build_system()	str | None — rendered system message
.build_string()	str — all messages concatenated
.build_last_user()	str | None — last user message
.token_count(model)	int — estimated token count
.missing_vars()	list[str] — unbound variable names

---

Template Engine

Uses {{double-brace}} syntax to distinguish templates from builder variables.

from prompt_builder_py import Template, TemplateRegistry

# Basic substitution with optional default
t = Template("Hello, {{name}}! You have {{count|0}} messages.")
print(t.render(name="Alice", count=3))
# Hello, Alice! You have 3 messages.

print(t.render(name="Bob"))
# Hello, Bob! You have 0 messages.

Conditionals

t = Template(
    "{{#if premium}}You have unlimited access.{{/if}}"
    "{{#unless premium}}Upgrade to unlock all features.{{/unless}}"
)
print(t.render(premium=True))   # You have unlimited access.
print(t.render(premium=False))  # Upgrade to unlock all features.

Loops

t = Template(
    "Languages:\n{{#each langs}}- {{@index}}. {{.}}\n{{/each}}"
)
print(t.render(langs=["Python", "Go", "Rust"]))
# Languages:
# - 0. Python
# - 1. Go
# - 2. Rust

# Loop over dicts
t2 = Template("{{#each users}}{{name}} ({{role}})\n{{/each}}")
print(t2.render(users=[
    {"name": "Alice", "role": "admin"},
    {"name": "Bob",   "role": "viewer"},
]))

Partials and registry

TemplateRegistry.register("footer", "\n---\nprompt-builder-py")
TemplateRegistry.register(
    "reply",
    "Answer in {{lang|English}}:\n{{question}}\n{{> footer}}"
)

result = TemplateRegistry.render("reply", question="What is Python?", lang="French")

Introspection

t = Template("Hello {{name}}, your score is {{score|N/A}}.")
print(t.required_vars())  # ["name"]
print(t.optional_vars())  # ["score"]
print(t.validate({"score": 42}))  # ["name"]  ← missing

---

Chat History

from prompt_builder_py import ChatHistory

history = ChatHistory(max_tokens=4096, model="gpt-4o")
history.add_system("You are a concise assistant.")
history.add_user("What is Python?")
history.add_assistant("A high-level, interpreted programming language.")
history.add_user("And Rust?")

print(len(history))          # 4
print(history.token_count()) # ≈ estimated tokens

# Export
print(history.to_markdown())
json_str = history.to_json()

# Restore from JSON
history2 = ChatHistory.from_json(json_str, max_tokens=4096)

Truncation strategies

# oldest (default): drop oldest non-system messages
history = ChatHistory(max_messages=10, truncation_strategy="oldest")

# pairs: drop oldest user/assistant exchange pairs
history = ChatHistory(max_tokens=2048, truncation_strategy="pairs")

# middle: keep first and last messages, drop the middle
history = ChatHistory(max_tokens=2048, truncation_strategy="middle")

# Manual truncation
history.truncate(strategy="oldest")

Search

results = history.search("Python")
for msg in results:
    print(msg.role, msg.content[:50])

---

Token Counting

from prompt_builder_py import TokenCounter, TokenBudget

counter = TokenCounter(model="gpt-4o")
print(counter.count("Hello, world!"))           # ≈ 4
print(counter.count_messages([
    {"role": "user", "content": "Hello there"}
]))                                              # ≈ 11

# Truncate a long string to fit in a budget
short = counter.truncate_to_budget(long_text, budget=1000)

# Split into chunks
chunks = counter.split_by_budget(long_text, budget=500)

budget = TokenBudget(max_tokens=8192, model="gpt-4o")
budget.consume("System prompt text here")
budget.consume_messages(messages)
print(budget.remaining)    # tokens left
print(budget.usage_pct)    # e.g. 23.4
print(budget.is_exhausted) # False

---

Provider Formatters

OpenAI

from prompt_builder_py import OpenAIFormatter

payload = OpenAIFormatter(model="gpt-4o").format(
    messages,
    max_tokens=1024,
    temperature=0.7,
    response_format={"type": "json_object"},
)

Anthropic

from prompt_builder_py import AnthropicFormatter

payload = AnthropicFormatter(model="claude-sonnet-4-5").format(
    messages,
    max_tokens=2048,
    system="Override system prompt here",
)

Cohere

from prompt_builder_py import CohereFormatter

payload = CohereFormatter(model="command-r-plus").format(messages, max_tokens=512)

Mistral

from prompt_builder_py import MistralFormatter

payload = MistralFormatter(model="mistral-large-latest").format(messages, max_tokens=1024)

Plain text

from prompt_builder_py import TextFormatter

payload = TextFormatter().format(messages, separator="\n\n", role_prefix=True)
print(payload["text"])

---

Validation

from prompt_builder_py import validate_messages, MessageValidator, ValidationResult

# Simple helper — raises ValidationError on failure
validate_messages(messages, provider="anthropic", max_tokens=4096)

# Full control
result: ValidationResult = MessageValidator().validate(
    messages,
    provider="openai",
    max_tokens=8192,
    strict_alternation=True,
    require_user_first=True,
)
print(result.is_valid)   # True / False
print(result.errors)     # list of error strings
print(result.warnings)   # list of warning strings
result.raise_if_invalid()

---

Complete Example

from prompt_builder_py import (
    PromptBuilder,
    ChatHistory,
    Template,
    TemplateRegistry,
    AnthropicFormatter,
    TokenBudget,
    validate_messages,
)

# 1. Register reusable templates
TemplateRegistry.register(
    "code_review",
    "Review the following {{lang|Python}} code for correctness, "
    "performance, and security.\\n\\n```{{lang|python}}\\n{{code}}\\n```"
    "{{#if focus}}\\n\\nFocus on: {{focus}}{{/if}}"
)

# 2. Build the prompt
builder = (
    PromptBuilder()
    .system("You are a senior software engineer conducting a thorough code review.")
    .user(TemplateRegistry.render(
        "code_review",
        lang="Python",
        code="def add(a, b): return a + b",
        focus="type safety",
    ))
    .strict()
)

# 3. Track token usage
budget = TokenBudget(max_tokens=4096)
messages = builder.build()
budget.consume_messages(messages)
print(f"Tokens used: {budget.used} / {budget.max_tokens} ({budget.usage_pct}%)")

# 4. Validate before sending
validate_messages(messages, provider="anthropic", max_tokens=4096)

# 5. Format for Anthropic
payload = AnthropicFormatter().format(messages, max_tokens=1024)

# 6. Persist the conversation
history = ChatHistory(max_tokens=4096)
history.extend(messages)
history.add_assistant("The code looks correct. Consider adding type hints…")
print(history.to_markdown())

---

API Reference

PromptBuilder

Method	Description
.system(content, **meta)	Append a system message
.user(content, **meta)	Append a user message
.assistant(content, **meta)	Append an assistant message
.add(role, content, **meta)	Append a message with any role
.extend(messages)	Append a list of dicts
.var(name, value)	Bind one variable
.vars(mapping, **kw)	Bind multiple variables
.unset(*names)	Remove variable bindings
.strict(enabled)	Toggle strict variable mode
.copy()	Deep-copy the builder
.reset()	Clear all state
.build()	Build all messages
.build_chat()	Build non-system messages
.build_system()	Get rendered system prompt
.build_string(sep)	Build as single string
.token_count(model)	Estimate token count
.missing_vars()	List unbound variables

Template

Method	Description
.render(context, **kw)	Render with context
.required_vars()	Variables with no default
.optional_vars()	Variables with defaults
.all_vars()	All referenced variables
.validate(context)	List missing required vars

ChatHistory

Method	Description
.add_system(content)	Set the system message
.add_user(content)	Append a user message
.add_assistant(content)	Append an assistant message
.extend(messages)	Append a list of dicts
.to_messages()	Export as list[dict]
.to_json()	Serialize to JSON string
.from_json(data)	Deserialize from JSON
.to_markdown()	Format as Markdown
.search(query)	Find messages by content
.token_count()	Estimate token count
.truncate(strategy)	Trim to configured limits

---

License

MIT © Vladyslav Zaiets