Deep Research Skills

Define objects and fields, then investigate in batches

Weizhena’s five research Skills turn a topic into an editable item list and field definitions, investigate items into JSON, then assemble a separate Markdown report. Useful for comparisons and surveys that need consistent questions and traceable gaps.

By Weizhena

Setup & first task View on GitHub

Hosts listed by the author

Claude Code
Codex
OpenCode

GitHub stars: 2,441
Checked: 2026-10-30
Categories: Writing & Research Content & Knowledge

Useful when the comparison criteria matter as much as the answer

Use it when each object needs the same set of questions: a tool shortlist, a paper survey or a competitor inventory. You retain item-level data and can add criteria before continuing. Approve a small outline and inspect the first batch of JSON before increasing the scale.

Pick a task and get started

These three tasks form a sequence: approve the outline, authorize investigation, then assemble the report. You can stop after the first task; later tasks need the actual files from the earlier stages.

Choose your current stage

Build an editable shortlist firstApprove a batch and inspect the evidenceTurn checked JSON into a readable report

Start with three candidates and stop after the outline. This is an editorial research exercise, not a prewritten evaluation of these tools.

Install the complete Skills and research agent
Choose one language, copy all five Skills plus the agent and modules, and prepare the Python dependency. The pinned Claude Code path below uses English; see the client-specific notes before using Codex or OpenCode.
Read the author’s setup guide
Claude Code in English: install the full set, then open a workspace
First check whether ~/.claude/skills already contains research, research-add-items, research-add-fields, research-deep or research-report, or whether the research agent has local changes to preserve. Copying updates files with the same names. These commands pin the reviewed commit, create a Python environment and open research-workspace beside the repository. Check the five entry points and web-search-agent before using the first task. Install only one language: replace research-en with research-zh for Chinese. In future sessions, activate this repository’s .venv before entering the workspace and launching the client.
```
git clone https://github.com/Weizhena/Deep-Research-skills.git deep-research-skills
cd deep-research-skills
git checkout 6ce38f60e3f8b22502c29873f96503a4e0c5addb
python3 -m venv .venv
source .venv/bin/activate
python -m pip install pyyaml
mkdir -p ~/.claude/skills ~/.claude/agents
cp -r skills/research-en/* ~/.claude/skills/
cp agents/web-search-agent.md ~/.claude/agents/
cp -r agents/web-search-modules ~/.claude/agents/
mkdir -p ../research-workspace
cd ../research-workspace
claude
```
Read the author’s installation steps
Give your AI this first task
Examples use the English Claude Code Skills. Open the client in the research workspace and use real paths. Deliverables are expected outputs; no third-party research workflow was executed for this page.
/research Browser regression testing for a small team I installed the English Skills, web-search-agent and strategy modules using this page, and opened Claude Code in an empty research-workspace. First check that the five Skills, research subagent, web search and a Python interpreter that can import yaml are available. Identify anything missing. Start with Playwright, Cypress and WebdriverIO. We maintain one web application and want to compare running tests in our existing CI and investigating failures. Do not install these tools or run benchmarks. Get the current date first. Use sources from the last 12 months, with current product behavior checked against official documentation as of the search date. Show proposed objects and fields first; ask before adding candidates. Put the time window and purpose into item descriptions or field definitions so downstream agents receive them. Create browser-testing/outline.yaml and fields.yaml. Suggested fields: name, official_url, checked_on, core_license, ci_execution, debugging_artifacts, setup_requirements and sources. In sources, associate key fields with original URLs, supporting content and access dates; distinguish facts from inference. Use fields → category name → a list of dictionaries containing name, description and detail_level. Do not add required markers for this exercise; use flat JSON results later. Confirm batch_size=2 and items_per_agent=1. Set execution.output_dir to the absolute path of browser-testing/results. Show the full paths, item count and field definitions for my approval, then stop at the outline.
Check what you receive
- Expect browser-testing/outline.yaml for objects and execution settings, plus fields.yaml for field definitions and detail levels. Per-item findings do not exist yet.
- Check candidates, time scope, source fields, two parallel agents and one item per agent. Confirm the output directory belongs to this topic and the field schema matches the validator.
- Use /research-add-items or /research-add-fields to revise and approve the plan before starting the deep investigation.
Read the author’s notes for this task

Before you start

A client with subagents and web access: The copyable path on this page uses Claude Code. The author lists direct /research-style invocation for 2.1.0+. Reading, writing, Bash, web search and subagent tools must be available. The research agent defaults to model: opus, so your account needs access and sufficient usage allowance.
All five Skills plus the search modules: Copy the complete chosen language directory, then install web-search-agent.md and web-search-modules separately. The five modules describe strategies for papers, general web research, GitHub, Chinese communities and Q&A; they do not configure accounts or grant database access.
Git, Python and PyYAML: The commands below target a macOS/Linux shell. The validator imports yaml and the templates invoke python. A virtual environment keeps installation and execution on the same interpreter; activate it again before launching Claude Code from a new terminal.
A separate workspace and a modest first batch: Run from the parent of the topic directories: later Skills look for */outline.yaml or */fields.yaml. Start with one topic and an absolute output_dir. Keep the initial item count and concurrency small; search and multiple agents consume your client’s model allowance.

Who it’s for

Professionals · Students & researchers · Developers

Know the boundaries

The client supplies search, parallel agents, file access and Python execution. Skills and strategy modules do not include a search service, website accounts or a guaranteed completion time.
Approval before another batch and validation before completion are author instructions. Check actual agent behavior and client permissions during use; written requirements are not unbypassable system controls.
Resume instructions describe skipping completed JSON. Added fields, stale findings and damaged files need another review; automatic refresh is not defined.
The MIT license permits use and modification while requiring the copyright and permission notice in copies or substantial portions. Model services, search services and retrieved material retain their own terms.

Keep outline.yaml, fields.yaml, raw JSON, the generated converter and the report together. Deliver unknown and unfinished items alongside the report so the next research pass has a clear starting point.

More about this skill

This collection separates research into five explicit entry points. /research first proposes objects and fields from model knowledge, then asks a research agent to supplement them through web search. After your review, outline.yaml records the objects and batch settings, while fields.yaml defines what to learn about each object and how much detail to collect.

The plan remains editable. /research-add-items extends the object list, and /research-add-fields adds comparison criteria. Once approved, /research-deep gives agents the shared field definitions, writes one JSON result per object and runs a field-coverage check. batch_size means parallel agents; items_per_agent controls how many objects each handles. The instructions call for approval between batches and describe resuming by skipping completed JSON files.

/research-report is a separate step. It generates generate_report.py from the results, runs that script and writes report.md with an anchored table of contents and grouped details. The contents can include short metrics present in the data. Unknown values are omitted from the body, which makes raw JSON and an explicit gap list valuable companions to the report. A synthesis of published material is not evidence that the products were tested.

The repository includes Chinese and English directories for Claude Code/OpenCode and separate Codex directories. This page uses the complete Claude Code setup for a first run and explains issues to check in the current Codex/OpenCode configuration. It is a different project from 199-biotechnologies’ Claude Deep Research: this collection centers on an editable object-and-field matrix; the other project centers on evidence records, claim verification and a multi-stage analytical report.

Sources and verification

Sources checked 2026-10-30

Documentation, Skills, agents, modules and validators checked at commit 6ce38f60e3f8b22502c29873f96503a4e0c5addb. This page reports source review and static checks only.

Author documentation Original skill

Another way to organize research

Compare Claude Deep Research when the central task is organizing evidence around claims and a longer report. Installation and invocation are separate.

Deep ResearchScope a question, keep sources and evidence, and write a cited report with counterevidence and limits.Obsidian SkillsSix skills for Obsidian notes, Canvas, Bases, app commands, web extraction and template batches.SwarmVaultOrganize documents and code into on-disk Markdown, a knowledge graph and a search index.Renwei WritingLightly edit personal writing and transcribed speech while keeping perspective and cadence.

Useful when the comparison criteria matter as much as the answer

Pick a task and get started

Install the complete Skills and research agent

Claude Code in English: install the full set, then open a workspace

Give your AI this first task

Check what you receive

Sources and verification

Another way to organize research