Skip to content

jayhjenkins/research-agent

Repository files navigation

Researcher Agent System

A skills-first, composable research orchestration framework for Claude Code. This system enables deep, persistent, and verifiable research through structured workflows, subagent coordination, and progressive disclosure.

Architecture Overview

researcher-agent/
├── skills/           # Primary instruction set (skills-first design)
│   ├── meta/         # Bootstrap and skill discovery
│   ├── project-setup/
│   ├── clarify-and-scope/
│   ├── prompt-rewrite/
│   ├── research-plan/
│   ├── web-research/
│   ├── data-extraction/
│   ├── analysis-and-modeling/
│   ├── synthesis-writing/
│   ├── quality-assurance/
│   ├── iteration-and-extension/
│   └── tooling-onboarding/
├── agents/           # Subagent definitions
├── commands/         # User-facing workflow triggers
├── templates/        # Report and artifact templates
├── lib/              # Helper scripts and utilities
├── docs/             # Documentation and examples
└── research_projects/ # Persistent project folders

Quick Start

Starting a New Research Project

Step 1: Invoke the research command

/research:new <your research question or topic>

Step 2: Answer clarifying questions (if any) The system will ask targeted follow-up questions only when necessary to scope the research properly.

Step 3: Review and approve the research plan The planner subagent generates a fit-for-purpose plan. Review it and confirm to proceed.

That's it. The system handles the rest and delivers:

  • A clean narrative report at outputs/report.md
  • An evidence table at the end of the report
  • A research effort summary at outputs/effort_summary.md
  • All sources and working artifacts preserved in the project folder

Continuing Existing Research

/research:continue <project-path>

Adding Supplemental Research

/research:supplement <project-path> <new question or angle>

Core Concepts

Skills-First Design

Skills are the primary system primitive. Every capability is expressed as a skill with:

  • SKILL.md — YAML frontmatter (name, description, triggers, inputs, outputs) + concise instructions
  • Supporting files — Checklists, rubrics, examples loaded only when needed (progressive disclosure)

Mandatory usage rule: If a relevant skill exists, the agent MUST use it.

Subagent Coordination

Research is orchestrated through specialized subagents:

Subagent Role
Planner Scopes research, creates plans, assigns tasks
Researcher Web search, source gathering, citation capture
Analyst Calculations, modeling, data processing
Skeptic Verification, gap-finding, contradiction detection
Writer Narrative synthesis, report drafting
Editor Structure, clarity, formatting, final polish

Subagents work in parallel where possible, writing outputs to designated folders.

Project Persistence

Every research request creates a persistent folder:

research_projects/<subject>/<YYYY-MM-DD>__<request_slug>/
├── inputs/           # Original prompt, clarifications, constraints
├── plans/            # Research plans, query strategies, subagent briefs
├── work/             # Notes, drafts, calculations, code experiments
│   └── code/         # Scripts for deterministic processing
├── sources/          # Captured excerpts, downloads, raw data
├── outputs/          # Final deliverables
│   ├── report.md     # Main narrative report
│   ├── effort_summary.md
│   └── appendix.md   # (optional)
└── meta/
    ├── project.yaml  # Scope, status, key questions
    └── changelog.md  # Iteration history

Research Workflow

Phase 0: Intake

  • Parse user request
  • Determine if clarification is needed
  • Ask follow-up questions BEFORE starting research

Phase 1: Setup & Planning

  • Create project folder structure
  • Generate expanded research brief
  • Produce adaptive research plan

Phase 2: Research Orchestration

  • Planner refines and assigns tasks
  • Researcher gathers and captures sources
  • Analyst processes data and builds models
  • Skeptic audits claims and identifies gaps
  • All artifacts persist to project folder

Phase 3: Synthesis

  • Writer produces narrative report
  • Evidence table appended at end
  • Clean structure following template

Phase 4: Quality Assurance

  • Verify claim-evidence mapping
  • Check uncertainty labeling
  • Confirm open questions listed
  • Validate deliverable meets request

Phase 5: Delivery

  • Print medium-depth TL;DR to user
  • Point to project folder and report location
  • Generate effort summary

Quality Standards

Evidence Requirements

  • Every key claim must appear in the Evidence Table
  • Prefer primary sources; triangulate when possible
  • Explicitly label uncertainty and confidence levels

Report Structure

  • Clean narrative without inline evidence clutter
  • Evidence Table at END of report
  • TL;DR at appropriate depth (not too short, not full report)

Uncertainty Handling

  • Surface unresolved contradictions
  • Explain why uncertainty remains
  • List open questions for future research

Extending the System

Adding New Skills

See skills/meta/skill-creation/ for the skill authoring guide.

Adding Tools/MCP Servers

See skills/tooling-onboarding/ for the tool integration guide. The system is designed for future MCP integration without refactoring.

Customizing Templates

Modify files in templates/ to adjust report structure, evidence table format, or project metadata schema.


File Reference

Key Files

  • skills/meta/using-skills/SKILL.md — Bootstrap skill (loaded at session start)
  • templates/report_outline.md — Report structure template
  • templates/evidence_table.md — Evidence table specification
  • agents/*.md — Subagent instruction files

Commands

  • commands/new_research.md — Start new research project
  • commands/continue.md — Continue existing project
  • commands/supplement.md — Add supplemental research

Example Session

See docs/worked_example.md for a complete walkthrough showing:

  1. User request
  2. Clarification exchange
  3. Generated plan
  4. Project folder structure
  5. Final report excerpt

Design Principles

  1. Skills as primitives — Capabilities are skills, not ad-hoc prompts
  2. Progressive disclosure — Load details only when needed
  3. Persistent artifacts — Everything is written to disk
  4. Observable process — User can inspect any intermediate state
  5. Composable subagents — Specialized roles with clear interfaces
  6. Adaptive planning — Plans fit the request, not rigid templates
  7. Quality gates — Verification is built-in, not optional
  8. Future-proof — MCP/tool integration ready without refactoring

About

General Researching Agent

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •