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

Forum topic · 小凯 · 2026-05-27

Summary

Deep-Research-skills is an MIT-licensed, open-source skill library by Weizhena that turns LLM coding assistants (Claude Code 2.1.0+, OpenCode, Codex) into structured, human-in-the-loop research pipelines. Inspired by the RhinoInsight paper (arXiv:2511.18743), it replaces the linear plan→search→write→report flow with a two-stage design—outline generation and deep investigation—gated by four explicit human confirmation nodes and persisted via JSON to support breakpoint-resume. The framework ships five slash-command skills (/research, /research-add-items, /research-add-fields, /research-deep, /research-report), five pluggable search-strategy modules (academic-papers, chinese-tech, github-debug, stackoverflow, general-web), and outputs validated against outline.yaml and fields.yaml through a three-layer quality gate. The OpenClaw adaptation wraps these as a /deep-flow entry, mapping WebSearch/WebFetch/Task to kimi_search/kimi_fetch/sessions_spawn, with proven deployment on 2026-05-27.

Overview

Deep-Research-skills (https://github.com/Weizhena/Deep-Research-skills) is an MIT-licensed, multi-platform research skill library authored by Weizhena (released 2025-12-29). It targets Claude Code, OpenCode, and Codex, and turns coding-assistant CLIs into structured research workflows with explicit human-in-the-loop control points, suitable for academic, technical, market, and due-diligence investigations.

Design Philosophy

The project engineers the ideas from the paper RhinoInsight: Improving Deep Research through Control Mechanisms for Model Behavior and Context (arXiv:2511.18743, Yu Lei, Shuzheng Si, Wei Wang et al., DeepLang AI + Tsinghua University, November 2025).

Three guiding principles:

1. Two-stage architecture — outline generation is extensible; deep investigation is traceable; the two stages connect via user confirmation. 2. Human-in-the-loop — every critical step requires user confirmation rather than a fully autonomous black-box run. 3. Breakpoint-resume — long investigations can pause and resume; JSON files act as durable intermediate state.

The workflow runs in three phases:

Phase 1 — /research: outline generation producing outline.yaml + fields.yaml
Phase 2 — /research-deep: parallel-agent deep investigation producing per-item JSON files
Phase 3 — /research-report: aggregation into report.md

Four human confirmation nodes: after outline (items + fields + time range), after each batch (continue? add items/fields?), after report (TOC format), and optional archive.

Skill System

| Command | Function | Trigger | |---|---|---| | /research | Generate outline with items and fields | Research kickoff | | /research-add-items | Add more research objects to existing outline | Missing dimensions discovered | | /research-add-fields | Add more field definitions to existing outline | Insufficient fields | | /research-deep | Parallel agents investigate each item | After outline confirmed | | /research-report | Render JSON results as markdown report | After investigation completes |

Deep phase auto-scans output_dir/*.json to skip completed items; JSON files are named {item_name_slug}.json.

Search Agents and Modules

Main agent: web-search-agent.md
5 strategy modules under web-search-modules/:
academic-papers — Google Scholar, arXiv, Semantic Scholar
chinese-tech — CSDN, Juejin, Zhihu, V2EX
github-debug — GitHub Issues, README, Releases
stackoverflow — Stack Overflow, Stack Exchange
general-web — Reddit, official docs, blogs, Hacker News

The agent reads the main config, then loads relevant strategy modules based on task type. Bilingual support (Chinese + English); Chinese-technical outputs are required.

Output Specifications

outline.yaml: topic, items (name/category/description), execution (batch_size, items_per_agent, output_dir)
fields.yaml: fields array with name/description/category/detail_level, plus uncertain placeholder array
detail_level: minimal → brief → detailed
category: basic info / technical features / performance metrics / milestones / business / competition & ecosystem / historical evolution / market positioning
JSON result: flat or nested (under category sub-dict); uncertain values marked "[不确定]" and listed in uncertain
report.md: anchor-linked TOC, skips fields containing [不确定]
validate_json.py: checks JSON coverage of all fields defined in fields.yaml and validates JSON syntax. Run: python validate_json.py -f {fields_path} -j {json_path}

Platform Adaptation

| Platform | Skill Path | Agent Path | Tooling | |---|---|---|---| | Claude Code 2.1.0+ | ~/.claude/skills/research-zh (Chinese) or research-en (English) | ~/.claude/agents/ | Native WebSearch + Task; requires pip install pyyaml | | OpenCode | Same path as Claude Code | ~/.config/opencode/agents/ | Requires OPENCODE_ENABLE_EXA=1 to enable websearch (otherwise only web fetch); Exa search API; requires pip install pyyaml | | Codex | ~/.codex/skills/research-codex-zh or research-codex-en | ~/.codex/agents/ | web-researcher.toml; requires multi_agent in ~/.codex/config.toml; auto-install via /scripts/install-codex.sh; requires pip install pyyaml |

Validation and Quality Control — Three Layers

1. Field coverage: validate_json.py flags missing fields 2. Human confirmation gates: pause and wait for user after each batch 3. Uncertain-value filter: report generator skips fields tagged [不确定] and entries in the uncertain array

Extensibility

/research-add-items and /research-add-fields dynamically expand an in-progress investigation.
Custom fields are freely defined via fields.yaml.
The five skills are decoupled through the file system and can be invoked individually or chained.

Installation (Quick Start)

1. pip install pyyaml 2. Clone the repo and copy the appropriate skill folder to the platform-specific path. 3. Copy agents/ to the platform-specific agents directory. 4. Trigger with /research <topic> → confirm outline → /research-deep → confirm per batch → /research-report.

A full topic investigation typically completes in 5–10 minutes.

OpenClaw Adaptation

Four adaptations were applied:

1. Path replacement: ~/.claude/skills/ → ~/.openclaw/skills/research-zh/deep-flow/ 2. Tool replacement: WebSearch → kimi_search; WebFetch → kimi_fetch; Task → sessions_spawn 3. Flow encapsulation: 5 skills wrapped into a single /deep-flow entry 4. allowed-tools adaptation: Read,Write,exec,kimi_search,kimi_fetch,feishu_ask_user_question

/deep-flow <topic> runs a 5-step closed loop: generate outline → confirm items/fields → run /research-deep in batches → generate report → optionally archive to zhichai.net.

Known limitations:

OpenClaw does not natively support Claude Code's /skills slash commands; trigger via natural language or alias.
sessions_spawn parallelism is limited by OpenClaw's concurrency policy.
kimi_search/kimi_fetch rate limits differ from Claude Code's WebSearch; adjust batch_size accordingly.
feishu_ask_user_question is only available on Feishu channels.
YAML frontmatter format differs slightly (e.g., aliases field).

Validation status: ✅ Deployed to ~/.openclaw/skills/research-zh/deep-flow/SKILL.md on 2026-05-27 21:18 and used to produce this meta-investigation.

Key Points

RhinoInsight (arXiv:2511.18743) is the academic backbone; the project engineeringizes its verifiable checklist and evidence audit into runtime file-based checkpoints.
Replacing a linear pipeline with a two-stage outline→deep loop, plus four explicit user confirmations, is the primary defense against context rot and accumulated error.
Five decoupled slash-command skills compose into a fully chained workflow but remain individually callable.
JSON-as-state enables true breakpoint-resume and cheap re-validation via validate_json.py.
Five search strategy modules cover academic, Chinese-tech, GitHub, Stack Overflow, and general web sources with bilingual support.
The OpenClaw adapter collapses five skills into /deep-flow and substitutes Kimi tools + Feishu interactive prompts, with verified deployment as of 2026-05-27.