SeanEllyJames/deep-research-skill
A prompt file that pushes AI agents to actually judge, not just summarize research
Most so-called 'deep research' done by AI agents is really just wide research: parallel searches across many sources, neatly sorted into a checklist that looks thorough but never actually takes a stance. This repository is a single drop-in Markdown skill file for tools like Claude Code and Cursor that forces the agent to question the question itself before searching, and to end with a conclusion someone in the field would actually argue about. It also lists six common failure modes agents fall into, each with a concrete warning sign.
What it does
- Contrasts default 'Wide Research' (parallel search, topic-sorted checklist) with 'Deep Research' (close reading of a few primary sources, argument chain from observation to judgment), arguing the former only produces well-organized but hollow output.
- Defines a four-phase workflow: Phase 0 interrogates hidden assumptions and the 'obvious but wrong' default answer before any search happens; Phase 1 requires close reading of 3 to 5 primary sources like earnings reports, founder blog posts, or call transcripts.
- Phase 2 asks for deep comparison of 2 to 3 cases while deliberately digging up failures and criticism; Phase 3 cross-validates sources, flagging single-source claims and treating contradictions as the most interesting findings.
- Phase 4 insists on prose over tables, one argument per section, honest labeling of unknowns, and a final self-check: rewrite if the core conclusion wouldn't spark any debate in the field.
- Documents six failure patterns with detection signals, including a real incident where recursive sub-agent spawning ballooned from 3 to 13 to 50+ to 149 agents, causing an unexpected API bill.
Why it matters
Anyone using AI agents to produce research or reports can borrow this as a concrete way to filter out polished-but-empty output and demand actual analytical judgment. It's also a practical warning against letting sub-agents recursively multiply and rack up runaway costs.
Terms in this repo
- Wide Research · Shallow research pattern: parallel searches across sources, sorted by topic
- Deep Research · Research pattern built on close reading of few primary sources and analyzing contradictions to reach a judgment
- SKILL.md · A Markdown file pasted into an AI agent to instruct a specific way of working
- subagent · A secondary AI agent that the main agent delegates sub-tasks to
- fan-out · When an agent keeps spawning more sub-agents branching from earlier ones
Repository description (English)
让AI Agent 做真正会思考，有判断的深度调研，而不是会摆信息的汇总报告。一个即插即用的 skill 文件。
Open on GitHubTrending repos
- vorssaint/vorssaint-utilsOne free menu bar app replaces a dozen paid Mac utilities
- Alishahryar1/free-claude-codeA local proxy that lets coding AI agents run on 49 free or cheap model providers instead of one paid service
- freestylefly/awesome-gpt-image-2A library of 532 reverse-engineered prompts that turn GPT-Image2 into a predictable image-making tool
- block/buzzAn open-source workspace where humans and AI agents chat, code, and review in the same rooms
- NousResearch/hermes-agentNous Research's Hermes is an AI agent that gets smarter the more you use it
- virgiliojr94/book-to-skillA tool that turns technical book PDFs into on-demand reference skills for AI coding agents
- VoltAgent/awesome-agent-skillsA single hub collecting over 1000 'how-to' manuals that make AI coding assistants act like experts
- anthropics/claude-plugins-communityA shared shelf where anyone's Claude add-ons get listed for install