deep-research

A systematic methodology for conducting thorough web research, meant to replace single, superficial searches with a structured, multi-angle process. The guidance is loaded before any research question or content-generation task, such as writing articles, building presentations, designing UI mockups, or producing multimedia, so that output is grounded in gathered evidence rather than general knowledge alone. The core principle is that a single search query is never enough, and that output quality depends directly on the depth and breadth of prior research.

The workflow moves through four phases. Broad exploration surveys the main topic and maps its dimensions, stakeholders, and viewpoints. Deep dive runs targeted queries for each dimension, tries multiple phrasings, and uses web_fetch to read the most authoritative sources in full rather than relying on snippets. Diversity and validation deliberately seek different information types: facts and data, real-world examples, expert opinions, trends, comparisons, and criticisms, to build a balanced picture. A synthesis check then confirms coverage across at least three to five angles before any content is written.

Practical guidance covers effective query patterns that add specificity and source hints, temporal awareness that reads the current date and chooses month-, day-, or year-level precision to match intent, and when to fetch full pages versus scanning results. It lists common mistakes to avoid, such as stopping after one or two searches, ignoring contradicting viewpoints, or generating content before research is complete. A quality bar defines when research is sufficient: when key facts, concrete examples, expert perspectives, current trends, and limitations can all be answered confidently.