Full automation is often a trap when you are trying to conduct high-stakes research.
I recently came across a fascinating post by a Reddit user named Electronic_Home5086 that challenges the current trend of “set it and forget it” AI agents. While everyone is rushing to build tools that do everything with one click, this expert argues, and proves, that keeping a human in the loop at specific decision points actually produces far superior results. The original poster based this workflow on principles from MIT research regarding recursive verification.
The core finding is simple but powerful: human oversight combined with AI speed beats full automation every single time when accuracy really matters.
The author developed a comprehensive 7-phase pattern to implement these principles. This isn’t just a list of tips; it is a structured, platform-agnostic workflow that you can use with ChatGPT, Claude, Perplexity, or even free tools. The creator tested this method for complex tasks like investment analysis and product strategy, finding that it consistently outperformed automated tools because it allows you to control exactly what information moves forward at each stage. It turns you from a passive observer into an active director of intelligence.
🧠 The Core Strategy: Recursive Verification
The fundamental concept the industry pro shares is that AI models struggle when you ask them to do too much at once. If you ask a model to “research the future of EV batteries and write a report,” it will likely hallucinate or miss nuance. The author’s solution is to break the process down into distinct phases where you verify the output before moving to the next step.
The creator outlines seven specific phases: Building a Map, Collecting Evidence, Deep Diving, Checking Quality, Writing the Report, Stress Testing, and Polishing. The genius here is that different phases use different types of AI models. You use reasoning models for planning and fast retrieval models for searching. By manually approving the output of one phase before feeding it into the next, you eliminate the “compounding errors” that plague fully automated agents. It might take 2 to 4 hours the first time you do it, but the author notes that once you learn the pattern, you can knock out complex, deep research in about 60 to 90 minutes.
🗺️ Insight 1: The “Project Manager” Phase is Critical
The first major insight from this expert is that you must never start researching without a plan. Most people skip straight to asking questions, but the author insists on a “Decomposition” phase first. In this stage, you are not asking the AI for answers; you are asking it to act as a project manager to break down your main objective into sub-questions.
The LinkedIn user suggests using a strong reasoning model for this, such as Claude Sonnet or the new o1 models. The goal is to generate 6 to 8 sub-questions that have clear dependencies. For example, the AI might identify that you cannot answer “market size” until you have first defined “target demographic.”
By using the specific prompt provided by the creator, you force the AI to list information requirements, authoritative source types, and search difficulty for each sub-question. This creates a roadmap. You, as the human, then review this map. If the map is wrong, the research will be wrong. This step alone prevents the AI from going down irrelevant rabbit holes later in the process.
🔎 Insight 2: Parallel Processing with Specific Constraints
Once the map is approved, the workflow moves to gathering evidence. The savvy professional points out that this is where you switch gears. You don’t need a slow, deep-thinking model here; you need a fast, accurate retrieval model like GPT-4o mini or Gemini. The author recommends running parallel searches, literally opening 3 or 4 separate threads, to tackle the sub-questions identified in the previous phase.
The creator’s prompt for this phase is highly specific. It demands full citations, key findings in bullet points, and a “credibility assessment” for every source. This is a game-changer because it forces the model to evaluate its own sources before presenting them to you. Instead of getting a generic summary, you get a list of facts with their origins attached.
Crucially, the expert advises that you (the human) must review these findings. You select which data points are valid and relevant. Only the verified information makes it into the “context” for the next step. This acts as a quality filter, ensuring that no hallucinations leak into the final writing phase.
🛡️ Insight 3: The Adversarial Stress Test
The final and perhaps most unique insight from this contributor is the “Stress Test” phase. Most users are happy once the AI writes a report, but the author argues that this is where the real work begins. Before you finalize anything, you should take your draft and feed it to a different AI model, preferably one known for logic and reasoning.
The goal is to ask this second AI to act as an adversary or a harsh critic. You ask it to find holes in the argument, identify weak evidence, or point out logical fallacies. This mimics the peer-review process in academia. By switching models (e.g., if you wrote with GPT-4, critique with Claude), you avoid the echo chamber effect where a model validates its own biases.
The original poster notes that this step often reveals contradictions that a tired human eye might miss. You then take these critiques and manually incorporate them into the final polish. This creates a level of depth and reliability that a single-shot prompt simply cannot achieve.
📝 Prompt of the Day: The Decomposition Planner
Here is the exact prompt the author designed for the first phase. It sets the foundation for the entire project. You should run this with a reasoning model.
Copy and paste this:
Research Objective: [Your main question – be specific]
Context:
– Purpose: [Why you need this – investment decision, product strategy, etc.]
– Scope: [Geographic region, time period, constraints, or ‘no constraints’]
– Depth needed: [Surface overview / Moderate / Deep analysis]
– Key stakeholders: [Who will use this, or ‘just for me’]Task: Create a comprehensive research plan
Break this into 6-8 sub-questions that together fully answer the objective. For each:
1. Specific information requirements (data, expert opinions, case studies, etc.)
2. Likely authoritative sources (academic papers, industry reports, government data, etc.)
3. Dependencies (which questions must be answered before others – be explicit)
4. Search difficulty (easy/moderate/hard)
5. Priority ranking (1-8, with 1 being highest)Output format:
– Numbered list of sub-questions
– For each: [Info needed] | [Source types] | [Dependencies] | [Difficulty] | [Priority]
– Final section: Recommended research sequence based on dependencies
If you want to master this workflow, I highly recommend looking at the original post for the full breakdown of all seven phases.
Check out the full post by Electronic_Home5086 on Reddit.
💡 FAQ & Troubleshooting
Which AI models should be used for the different phases?
For Phase 1 (Decomposition), use reasoning models such as Claude Sonnet, o1, or DeepSeek-R1 to create the research plan. For Phase 2 (Information Gathering), switch to fast retrieval models with web access, such as Gemini or GPT-4o mini.
Do I need a specific paid subscription for this workflow?
No. This pattern is platform-agnostic. While it works well with tools like Perplexity, it can be executed using free tiers of ChatGPT, Claude, or even manual search combined with the structured prompt logic.
How much time does the full 7-phase process require?
Expect the first attempt to take 2–4 hours as you learn the pattern. Once familiar with the workflow, complex research typically takes 60–90 minutes, while a condensed “quick version” can be completed in 30–45 minutes.
7-Phase Prompt Pattern for Deep Research (RLM-inspired, platform-agnostic)
byu/Electronic_Home5086 in