LangChain Prompt Debugger

LangChain PromptTemplate uses {variable_name} placeholders in prompt strings. At runtime, you call template.format(variable_name='value') to produce the final prompt. This makes prompts reusable, testable, and composable. ChatPromptTemplate works similarly but supports message roles (system, human, ai) for chat models.

Tokens are estimated at ~4 characters per token, which is a common approximation for English text with OpenAI and Anthropic models. Actual token counts depend on the tokenizer: GPT models use tiktoken (BPE), Claude uses a similar approach, and Gemini uses SentencePiece. For exact counts, use the Tokenizer Counter tool or the model's API.

PromptTemplate produces a single string prompt suitable for completion models. ChatPromptTemplate produces a list of messages with roles (SystemMessage, HumanMessage, AIMessage) for chat models like GPT-4 and Claude. Use ChatPromptTemplate for all modern chat-based LLMs — it gives the model better context about each message's role.

Prompt injection happens when user-controlled input contains instructions that override your system prompt. Mitigations include: (1) wrapping user input in delimiters like XML tags <user_input>...</user_input>, (2) explicitly instructing the model to ignore instructions within user input, (3) using separate system and user message roles in ChatPromptTemplate, and (4) validating/sanitizing inputs before injecting them.

GPT-3.5 Turbo: 16K tokens. GPT-4o: 128K tokens. Claude 3.5 Sonnet: 200K tokens. Gemini 1.5 Pro: 1M tokens. Exceeding these limits causes API errors. Always leave buffer room for the completion output (typically 1K–4K tokens). This tool highlights when your filled prompt approaches common model limits.

Use LangChain Expression Language (LCEL) to chain components: `chain = prompt | llm | StrOutputParser()`. You can pipe the output of one chain as input to another: `chain = prompt1 | llm | prompt2 | llm`. For complex workflows, use LangGraph for stateful, graph-based chains with branching and loops.