TL;DR

• To overcome the limitations of a single model. Multi-LLM strategies that combine various models according to the nature of the task are being emphasized.
• This is necessary to suppress information loss and hallucinations that occur during long conversations and to improve accuracy at each stage.
• Complex tasks should be divided into detailed steps. Workflows should be constructed by combining structured prompting techniques with the strengths of specific models.

Example: A researcher performs tasks with multiple browser windows open. In one window, they enter references to verify facts; in another, they search for logical flaws in a draft; and in a third, they refine sentences to enhance the readability of the final output.

Expectations that a single Artificial Intelligence (AI) model will solve all problems are diminishing. Users are now focusing on a "Multi-LLM Strategy" that breaks down tasks according to the unique strengths of each model and connects them into an optimal workflow. The ability to acknowledge technical limitations and utilize tools for their intended purposes has become a key indicator of productivity.

Current Status

As major AI models become differentiated based on their design goals and strengths, users are utilizing specialized areas for each model. According to OpenAI guidelines, it is difficult to maintain previous content during long conversations due to inherent context limits. To address this, methods such as including summaries of previous conversations in system messages or filtering only core information are used. Since a model's attention can be dispersed as the flow of conversation becomes complex, it is efficient to start a new conversation or use Retrieval-Augmented Generation (RAG) to dynamically search only for necessary information.

In the research and data analysis stage, models with strong source grounding characteristics are used. Tools like NotebookLM generate answers based on materials provided by the user, which is advantageous for suppressing hallucinations and increasing precision through interactive feedback. Conversely, in the final review stage of a document, the ability to provide structural feedback while maintaining the context of the original text is crucial. Models like Claude show strengths in understanding the logical flow of an entire document and performing detailed proofreading.

Changes are also appearing in prompt design. Rather than simple text listings, techniques using Markdown format or XML tags to clearly set logical boundaries within a prompt are being utilized. This prevents the model from confusing instructions, reference materials, and user inputs, thereby increasing the accuracy of the output.

Analysis

The core of an optimized Multi-LLM strategy lies in context management and the separation of specialized areas for each model. Even for models with long context windows, "Middle Loss"—where information located in the middle is missed as input data increases—can occur. This is why OpenAI advises breaking complex tasks into smaller subtasks. Dividing task units increases the information density that the model needs to focus on at each stage.

Such a strategy is also important from a risk management perspective. Placing a model specialized in source-based reasoning, such as NotebookLM, at the beginning can prevent the generation of groundless answers. On the other hand, when using GPT-series models, guides should be provided through prompt restructuring so that the model does not forget instructions. Receiving structural feedback through models like Claude in the final stage serves as a mechanism to ensure the quality of the result by mimicking the human editing process.

However, utilizing multiple models may lead to task fragmentation. Core context may be lost when transferring information between models, and the operational costs incurred from switching between multiple tools should also be considered. Therefore, a process of summary and filtering to remove noise before inputting the output of each stage into the next model is necessary.

Practical Application

Practitioners should first divide the task into three stages: "Information Collection and Verification," "Drafting," and "Structural Revision." In the information collection stage, upload source documents to NotebookLM to confirm facts. Then, when drafting, utilize GPT, but if a long conversation ensues, summarize previous content and inject it into the system message or use XML tags to specify the data range to prevent model deviation.

In the final revision stage, use Claude to check the logical connectivity between sentences and the tone and manner. At this time, request the "structural feedback" function to ensure the original intent is not compromised. If the output from a specific model is unsatisfactory, it may be more efficient to replace it with another model better suited for that subtask rather than simply modifying the prompt.

To-Do Today:

• Summarize key information from an ongoing long conversation, open a new chat window, and reset it as a system message.
• Introduce XML tags such as <instruction> and <context> to distinguish between instructions and data within a prompt.
• For tasks where fact-checking is critical, directly upload reliable documents as sources to induce answers instead of using external searches.

FAQ

Q: Is it often better to start over when a conversation becomes too long? A: It is recommended to summarize previous key conclusions and context to serve as the foundation for the next conversation. OpenAI suggests filtering unnecessary conversation history and leaving only the core to maintain information density.

Q: Do Markdown or XML tags actually affect model performance? A: Yes. Models recognize hierarchies and logical boundaries of information more clearly through structured text. This contributes to improving instruction following, especially when instructions are complex or large amounts of data are input together.

Q: Doesn't using multiple models cost more time and money? A: While the initial setup may take time, it reduces the post-hoc costs of correcting hallucinations or logical errors that occur in a single model. For professional tasks where accuracy is paramount, combining the strengths of each model is more advantageous in terms of overall productivity.

Conclusion

Multi-LLM optimization by workflow is a strategy for effective task execution. It requires the design capability to recognize the limits of a model's context maintenance and to deploy different functions—such as source grounding and structural feedback—in the right places.

In the future, "workflow orchestration"—the ability to weave models with different characteristics into an organic pipeline—will determine a user's competitiveness more than the intelligence of individual models themselves. Rather than relying on a specific model, one should experiment with and verify combinations of tools that suit the nature of each task.

References

• 🛡️ Prompt engineering - OpenAI API
• 🛡️ Prompt engineering - OpenAI