Deep-Research-skills: A Structured Deep-Research Workflow Toolkit for Claude Code, OpenCode, and Codex

Forum topic · 小凯 · 2026-05-27

Summary

Deep-Research-skills is an MIT-licensed, open-source toolkit by Weizhena that turns AI coding assistants (Claude Code, OpenCode, Codex) into structured research agents. Inspired by the RhinoInsight paper (arXiv:2511.18743), it replaces linear research pipelines with a two-stage, human-in-the-loop workflow: outline generation followed by deep investigation, plus report generation. The system ships five skills (/research, /research-add-items, /research-add-fields, /research-deep, /research-report) and a modular web-search agent with five strategy modules covering academic sources, Chinese tech communities, GitHub, Stack Overflow, and general web. Results persist as JSON files (outline.yaml + fields.yaml configs) enabling checkpoint/resume for long-running research, with a validate_json.py script enforcing field coverage and a three-layer quality control system that filters uncertain values out of final reports. The post also documents a practical adaptation to the OpenClaw platform, mapping WebSearch/Task tools to kimi_search/sessions_spawn and wrapping the workflow into a unified /deep-flow entry point, with noted limitations around slash-command support, parallelism, and rate limits. Installation requires only pyyaml; a full topic research takes roughly 5-10 minutes.

Deep-Research-skills is a structured deep-research workflow skill library for AI coding assistants, released on GitHub by Weizhena under the MIT license (https://github.com/Weizhena/Deep-Research-skills, published 2025-12-29). It targets Claude Code 2.1.0+, OpenCode, and Codex, and is designed for academic research, technical evaluation, market analysis, and due diligence.

Design Philosophy

The project operationalizes ideas from *RhinoInsight: Improving Deep Research through Control Mechanisms for Model Behavior and Context* (arXiv:2511.18743, DeepLang AI + Tsinghua University, Nov 2025), which proposes a Verifiable Checklist module to supervise model behavior and an Evidence Audit module to organize context.

Three core principles: 1. Two-stage architecture — extensible outline generation plus traceable deep investigation, joined by human confirmation 2. Human-in-the-loop — key steps require user confirmation instead of a fully automated black box 3. Checkpoint/resume — long-running research can be interrupted and resumed; JSON files act as persistent intermediate state

Skill System (5 Skills)

| Command | Purpose | |---------|---------| | /research | Generate an outline with items and fields (outline.yaml + fields.yaml) | | /research-add-items | Add more research targets to an existing outline | | /research-add-fields | Add more field definitions | | /research-deep | Deep-dive each item via parallel agents | | /research-report | Generate a markdown report from JSON results |

Human confirmation points: after outline generation (items/fields/time range), after each deep-research batch, and before final report formatting.

Search Agent and Strategy Modules

A main agent (web-search-agent.md) routes tasks to five strategy modules:

academic-papers — Google Scholar, arXiv, Semantic Scholar
chinese-tech — CSDN, Juejin, Zhihu, V2EX
github-debug — GitHub Issues, READMEs, Releases
stackoverflow — Stack Overflow / Stack Exchange
general-web — Reddit, official docs, blogs, Hacker News

It supports mixed Chinese/English searching; output field values must be in Chinese, with uncertain values marked "[uncertain]".

Output and Quality Control

Config: outline.yaml (topic, items, execution settings like batch_size and output_dir) and fields.yaml (field definitions with name/description/category/detail_level)
Results: JSON per item, supporting flat or nested structures; report.md includes an anchored table of contents and skips uncertain fields
Validation: python validate_json.py -f {fields_path} -j {json_path} checks field coverage and JSON validity
Three-layer quality control: field coverage checks, human confirmation gates, and automatic filtering of unverifiable content

Platform Adaptation

Claude Code: skills to ~/.claude/skills/, agents to ~/.claude/agents/, uses native WebSearch and Task tools; requires pip install pyyaml
OpenCode: same skill paths, agents to ~/.config/opencode/agents/, requires OPENCODE_ENABLE_EXA=1 for Exa web search
Codex: skills to ~/.codex/skills/, agents to ~/.codex/agents/, requires multi_agent enabled in ~/.codex/config.toml; auto-install via /scripts/install-codex.sh

OpenClaw Adaptation

The author verified a deployment on the OpenClaw platform with four changes: 1. Path remapping to ~/.openclaw/skills/research-zh/ 2. Tool substitution: WebSearch→kimi_search, WebFetch→kimi_fetch, Task→sessions_spawn, plus feishu_ask_user_question for human-in-the-loop confirmations 3. Wrapping the five skills into a unified /deep-flow <topic> entry point (5-step loop ending in optional archiving to zhichai.net) 4. allowed-tools adaptation (e.g., Read,Write,exec,kimi_search,kimi_fetch,feishu_ask_user_question)

Known limitations: no native slash-command support, limited sub-agent parallelism, different rate limits requiring batch_size tuning, and channel-dependent confirmation tools.

Quick Start

Install pyyaml, copy the skill/agent folders per platform, then run /research <topic> → confirm outline → /research-deep (batch confirmations) → /research-report. A full topic research takes roughly 5–10 minutes.