2 Why This Example Works.

This example applies the study's methodology effectively to tasks that demand accuracy and fact verification:

1. Co-planning for research: Asking for a "research plan" forces the LLM to structure its approach. It lets the user confirm the LLM understood the task and will look for information in the right places, instead of just stitching together the first facts it finds online.
2. Lower hallucination risk (Reliability): The "checkpoint" that requires sources (cite sources for every point) is a practical implementation of the "Answer Verification" and "Action Guards" principles. The user doesn't just trust the LLM; they get a chance to check facts at an intermediate stage. That is critical for topics where accuracy matters and helps fight made-up facts (hallucinations).
3. Building trust step by step: By breaking a complex task into manageable, verifiable steps, the user builds trust in the process. Instead of getting a big "black box" of text at the end and guessing what's true and what isn't, the user takes an active part in creating it and controls quality at every stage.

Key Aspects of the Research:

The study shows that fully autonomous LLM agents often fail at complex tasks. Instead of full autonomy, it proposes a human-in-the-loop model where the user and the agent work as a team. To do this, the authors built Magentic-UI, a system that implements co-planning, co-execution of tasks and human oversight.

Key Result: Involving a human in the process through these mechanisms significantly raises the task success rate compared to fully autonomous systems.

The Method in a Nutshell:

The core idea is to stop treating the LLM as a "black box" that should execute a complex instruction perfectly on the first try. Instead, you become the LLM's manager or partner.

For an everyday user, the approach based on the study's findings looks like this:

1. Co-planning: Don't hand the LLM the whole complex task at once. First ask it to draft a step-by-step action plan.
   • Example command: "Before you start writing, draft a detailed plan for how you'll do it. Show me the plan for approval."
2. Edit the plan: Review the proposed plan. If it's inaccurate, incomplete or heading the wrong way, fix it right there in the chat. You know the context and the end goal better.
   • Example command: "The plan is good, but in step 2 search the Microsoft website instead of Amazon. And drop step 4, we don't need it."
3. Controlled execution (Co-tasking & Action Guards): Tell it to execute the approved plan. For the most important steps, build "checkpoints" into the instruction: moments where the LLM must stop and ask for your permission or choice.
   • Example command: "Now execute the plan. When you get to step 3 (choosing a hotel), give me 3 options with prices and wait for my pick before continuing."
4. Save what works (Memory): If a conversation and prompt structure worked well for a certain type of task, save that template for future use.

This approach turns a one-shot prompt into a guided conversation where you steer the LLM, fix its mistakes on the fly and make sure the result matches your expectations.

Practical Applicability:

• Direct applicability: Very high. Any user can start applying co-planning in their chats with ChatGPT, Claude and other LLMs right away. Instead of one giant prompt, break the interaction into stages:

  1. "Draft a plan for [my task]"
  2. "Plan approved. Start on step 1"
  3. "Great. Before you do step 2, show me the options..." This carries the Magentic-UI logic straight into a regular chat.

• Conceptual value: Huge. The study gives users a new "mental model": an LLM is a powerful but inexperienced intern. You can't leave it unsupervised with a complex task. You need to guide it, give clear instructions, check intermediate results and correct course. That understanding dramatically improves how effectively you work with it.

• Adaptation potential: Maximum. The adaptation is simple: the user mentally takes on the role of the "Orchestrator" from Magentic-UI. They decide when to ask for a plan, when to approve it and when to step in. The whole complex dialogue that Magentic-UI automates, the user reproduces by hand in a regular chat, steering the LLM through a series of short, clear commands.

1 Practical Example:

You are an experienced social media manager. Your task is to create a one-week content plan for the Instagram account of a coffee shop called "Cozy Corner".

**STAGE 1: CO-PLANNING**

Your first and **only** task at this step is to propose a step-by-step plan for **HOW** you will build the content plan. The plan should cover target audience analysis, defining content pillars, and generating ideas for posts and stories.

**IMPORTANT:** Do not generate the content plan itself! Just show me your plan of action for approval. Wait for my further instructions.

(After the LLM proposes a plan, the user adjusts it and gives the next command)

Great, your plan of action is approved.

**STAGE 2: EXECUTING THE PLAN WITH A CHECKPOINT**

Now start executing the plan. When you reach the "generate post ideas" step, stop.

Give me 5-7 specific post ideas as a table with the columns: "Idea", "Content Pillar", "Format (post/reels)".

**IMPORTANT:** After you show the table of ideas, **wait for my choice**. Do not continue until I tell you which ideas I like.

2 Why It Works:

This prompt works because it simulates the key principles of the Magentic-UI study:

1. Decomposition and co-planning: Instead of asking for a finished content plan right away (which may not match the user's vision), we make the LLM first lay out the process of creating it. That surfaces and fixes logic errors at the earliest stage.
2. Human oversight (Action Guard): The command IMPORTANT: ... Wait for my further instructions creates an artificial "checkpoint". The LLM can't "run ahead" and churn out content based on ideas the user might not like. That saves time and makes the final result more relevant.
3. Guided dialogue: The prompt turns a monologue (one request) into a guided dialogue where the user acts as project lead and the LLM as the contractor who reports back at every key stage.

1 Another Practical Example

You are my personal research assistant. I need to prepare a short summary of how microplastics affect human health, written for non-specialists.

**STEP 1: RESEARCH PLAN**

First, develop and present your research plan. The plan should include:
- Key questions that need answers.
- Types of sources you plan to use (scientific papers, WHO reports, popular publications).
- The structure of the final document.

**Critically important:** Do not start searching for information or writing text. Just show me the plan for sign-off.

(The user gets the plan, approves it and gives the next command)

Plan signed off. It's excellent.

**STEP 2: GATHERING INFORMATION WITH VERIFICATION**

Now start executing the plan. Gather information for each item.

**Checkpoint:** When you find information on the item "Main ways microplastics enter the body", stop. Present that information as short bullet points and **cite sources** for every point.

I'll check them, and only after my approval will you move on to the next item in the plan.