research

name	research
title	How to Do Great Research
description	Uplevel your researching abilities and learn how to research properly.
allowed-tools	["agents.*","files.*"]

General flow of the research process:

1. Analyze the topic and formulate a set (3-5) of diverse search queries to gather information.

2. List available files and see if there's already some relevant information that could be used as the initial foundation for research.

3. Create a child task of type generate_text (using the standard tasks tool with slug set to a unique folder name like search_1, search_2, etc., and objective set to your query) for each query in parallel.

4. Organize the information into a coherent whole. Deduplicate and order.

5. Save your findings to a file called research-notes.md.

6. Return the relative path of the file.

Best practices:

Rule 1: Do not rely on parametric memory for research. Mind the training data gap. You are an LLM, and your parametric memory is likely stale. Though they may seem like the present to you, things like today's date or key events that you know about it are in the past. There is a gap between what is happening in the real world and what you believe to be true. To close this gap, spawn a child task of type generate_text with search grounding to get the actual information. Rely on current_date tag in the metadata to orient in time.

Rule 2: Spawn multiple search child tasks in parallel. Before starting research, decide on different facets of information that would be useful to have and start a separate child task of type generate_text with search grounding for each query in parallel, specifying unique slug sandbox folders (e.g. search_1, search_2). They will run in parallel. It's a double-win: parallel-running saves time and it gives you rich data to organize.

Rule 3: Don't make assumptions or theorize. Good research does not attempt to fill in the blanks.

Rule 4: Don't summarize, organize. When you have a large cache of research data, you will be tempted to shorten the final output into a summary. That defeats the purpose of the research. Instead, organize the information, arranging data into a coherent whole. Keep it all. Deduplicate and order. Create a foundation for someone to build on.