Optimal Approaches for Forcing Pseudo-Multi-Agent Execution
To ensure that each role (persona) completely executes its assigned tasks—such as information gathering, fact-checking, and proofreading—without skipping them, transitioning to the Gemini API or CLI-based agent workflow tools is technically the optimal solution, rather than simply updating a single System Instruction.
The following sections objectively explain the structural limitations of current chat interfaces (which rely on a single System Instruction) and provide concrete solutions.
1. Structural Limitations of System Instructions (Chat UI)
Even if a System Instruction strictly mandates executing “9 sequential steps,” the inherent architecture of Large Language Models (LLMs) often leads to skipped processes (shortcuts) for the following reasons:
- Single-Pass Generation Bias: LLMs attempt to generate the final output (the article) in a single processing pass. Consequently, they often skip internal loops like “launching search tools,” “receiving user feedback,” and “rewriting drafts,” instead outputting a highly probable “first draft” as the final result.
- Unreliable Tool Execution: Even with strong commands like “always perform a web search,” the LLM’s internal mechanisms may bypass external tools if it calculates that its internal knowledge (past training data) is sufficient to generate a plausible answer.
2. Recommended Solution: Building Physical Workflows via Gemini CLI / APIs
To forcefully activate all roles and completely eliminate hallucinations, it is necessary to physically separate and serialize the processes using the Gemini API, CLI, or external frameworks such as CrewAI, LangGraph, orDify.
Building a Pipeline (Example)
Define each role as an independent AI agent (or individual API call) using programming or CLI scripts, and execute tasks in a relay format:
- Researcher Function: Call the Gemini API with Google Search Grounding to forcefully execute a web search on the specified topic. Output and save only the “raw text” of the results.
- Fact-Checker Function: Feed the Researcher’s raw text and the topic into a new prompt. Instruct it to “verify facts using only this text as the source, and create a list filtering out lies and dead links.”
- Writer Function: Pass only the fact-checked data to the prompt to generate the initial draft.
- User Representative & Editor Function: Pass the draft to a new prompt (User Persona) to generate harsh critique. Finally, pass the critique and the initial draft to another prompt to generate the final article.
Under this approach, it is 100% impossible for the system to skip searches or slack off, because the next role cannot proceed until the previous role’s output is complete.
3. Alternative Solution: Updating the Current System Instructions (Step-by-Step)
If building an API/CLI environment is not feasible and maximizing execution rates within the current Chat UI is required, the System Instructions must be updated from “outputting the article in one go” to an interactive step-by-step process.
Direction for Rule Updates (Prompt Segmentation)
Instead of forcing the AI to simulate everything internally, rewrite the rules to demand output for each process, moving to the next step only after user approval.
- Step 1: When given a topic, do not write the article. Output only the search result summary and the article outline, then stop.
- Step 2: Once the user approves with “OK,” output bullet points of the User Representative’s critique and the Fact-Checker’s results based on the outline, then stop.
- Step 3: Upon receiving the final “OK,” output the final article text formatted for WordPress (H2/H3, inline links, excerpts, slugs, etc.).
By forcing the AI to stop and output its thought process and tool execution at each step, the “shortcuts” typical of single-pass generation can be drastically reduced.
