/orion-help - Orion Documentation

Display help and documentation for the Orion AI Research Automation System.

Usage

/orion-help [topic]

Topics

• overview - What is Orion?
• workflow - Complete research workflow
• commands - Available commands
• skills - Tree of Skills system
• plan-mode - How Plan Mode works
• troubleshooting - Common issues

Process

When user invokes /orion-help, display the relevant documentation.

Default (no topic)

╔══════════════════════════════════════════════════════════════╗
║              ORION - AI Research Automation                   ║
╠══════════════════════════════════════════════════════════════╣
║                                                               ║
║  Automate AI research: papers → knowledge → hypotheses → test ║
║                                                               ║
╚══════════════════════════════════════════════════════════════╝

## Commands

| Command                 | Purpose                           |
|-------------------------|-----------------------------------|
| /orion-init            | Initialize project structure       |
| /knowledge             | View accumulated knowledge         |
| /literature-review     | Read papers, extract skills        |
| /benchmark-setup       | Download/study benchmarks          |
| /orion-setup           | Clone and setup codebase           |
| /hypothesis-generation | Generate hypotheses → PRD → JSON   |
| /orion-help            | This documentation                 |

## Quick Start

```bash
/orion-init "my-project"
# Place papers in research/papers/
/literature-review research/papers
/knowledge
/benchmark-setup
/hypothesis-generation --target 75.0
/ralph-loop "Test hypotheses..."

For details: /orion-help <topic>

### Topic: workflow

Orion Research Workflow

Phase 1: Initialize

/orion-init "project-name"

Creates: research/ directory with skills structure

Phase 2: Literature Review

# Place PDFs in research/papers/
/literature-review research/papers

Extracts: methodology, experiments, metrics skills Uses: Plan Mode (creates plan, gets approval, executes)

Phase 3: View Knowledge (anytime)

/knowledge
/knowledge literature
/knowledge benchmarks

Shows: Accumulated knowledge summaries

Phase 4: Benchmark Setup

/benchmark-setup

Downloads: Benchmark data (with permission) Creates: Benchmark skills and evaluation helpers

Phase 5: Codebase Setup (if needed)

/orion-setup "https://github.com/author/paper-code"

Clones: Repository Verifies: Baseline reproduction

Phase 6: Hypothesis Generation

/hypothesis-generation --target 75.0 --benchmark VideoMME

Generates: Hypotheses from accumulated knowledge Creates: research-prd.md and research-prd.json

Phase 7: Experimentation

/ralph-loop "Test hypotheses in research/research-prd.json..."

Tests: Each hypothesis iteratively Documents: Learnings in skills/learned/

Flow Diagram

Papers → /literature-review → Skills
                ↓
         /benchmark-setup → More Skills
                ↓
        /hypothesis-generation → PRD + JSON
                ↓
           /ralph-loop → Results + Learnings

### Topic: commands

Orion Commands

/orion-init <project-name>

Initialize a new Orion research project.

Creates:

• research/orion.json (project state)
• research/skills/ (Tree of Skills)
• research/papers/ (place papers here)
• research/benchmarks/ (downloaded data)
• research/codebase/ (cloned repos)

Example:

/orion-init "video-understanding"

---

/knowledge [category]

View accumulated knowledge.

Categories:

• all (default) - Overview of all knowledge
• literature - Paper insights
• domain - User expertise
• benchmarks - Benchmark info
• learned - Experiment learnings
• status - Project progress

Example:

/knowledge
/knowledge literature

---

/literature-review <folder>

Read papers and extract skills. Uses Plan Mode.

Extracts per paper:

• Methodology (core technique, architecture)
• Experiments (benchmarks used, setup)
• Metrics (evaluation methods, results)

Example:

/literature-review research/papers

---

/benchmark-setup [name]

Setup evaluation benchmarks. Uses Plan Mode.

Does:

• Identifies benchmarks from literature
• Downloads (with permission)
• Creates evaluation helpers
• Documents in skills

Example:

/benchmark-setup
/benchmark-setup VideoMME

---

/orion-setup <repo-url>

Clone and setup codebase. Uses Plan Mode.

Does:

• Clones repository
• Sets up environment
• Verifies baseline
• Creates codebase skill

Example:

/orion-setup "https://github.com/author/paper-code"

---

/hypothesis-generation [--target <score>] [--benchmark <name>]

Generate hypotheses from knowledge. Uses Plan Mode.

Does:

• Synthesizes all accumulated knowledge
• Proposes hypotheses (user curates)
• Creates research-prd.md
• Creates research-prd.json for Ralph

Example:

/hypothesis-generation --target 75.0 --benchmark VideoMME

---

/orion-help [topic]

Display documentation (this command).

### Topic: skills

Tree of Skills

Orion organizes knowledge hierarchically to preserve context.

Structure

research/skills/
├── literature/              # From /literature-review
│   ├── _overview.md         # Summary (always in context)
│   ├── paper_001/
│   │   ├── summary.md       # Paper overview
│   │   ├── methodology.md   # Core technique
│   │   ├── experiments.md   # Benchmarks, setup
│   │   └── metrics.md       # Evaluation methods
│   └── paper_002/
│
├── domain/                  # User expertise
│   ├── _overview.md
│   └── [custom-skill].md
│
├── benchmarks/              # From /benchmark-setup
│   ├── _overview.md
│   └── [benchmark-name].md
│
└── learned/                 # From experiments
    ├── _overview.md
    └── h001_what_we_learned.md

Key Principle: Hierarchical Loading

• Default: Only _overview.md files are loaded
• On demand: Specific skills loaded when needed
• Prevents: Context blowup from full papers

Skill Categories

Literature Skills

Extracted from papers:

• Methodologies (how they did it)
• Experiments (what they tested on)
• Metrics (how they measured)

Domain Skills

User-provided:

• Expertise and intuitions
• Constraints (compute, time, data)
• Custom techniques

Benchmark Skills

From setup:

• Data format and structure
• Evaluation protocol
• Subset for quick testing

Learned Skills

From experiments:

• What worked and didn't
• Key insights
• Patterns discovered

### Topic: plan-mode

Plan Mode

Most Orion commands use Plan Mode for user control.

How It Works

1. Claude creates a plan

   • Analyzes what needs to be done
   • Lists specific steps
   • Identifies decisions needed

2. User reviews

   • See exactly what will happen
   • Approve, modify, or reject

3. Execution

   • After approval, Claude executes
   • Progress visible throughout

Commands Using Plan Mode

Command	What's Planned
/literature-review	Papers to read, skills to extract
/benchmark-setup	What to download, where to store
/orion-setup	Clone location, environment setup
/hypothesis-generation	Hypotheses, priority order

Commands NOT Using Plan Mode

Command	Why
/orion-init	Simple directory creation
/knowledge	Read-only viewing
/orion-help	Documentation display

Benefits

• Transparency: Know what will happen
• Control: Approve before execution
• Safety: No surprises
• Iteration: Modify plans before running

### Topic: troubleshooting

Troubleshooting

Common Issues

"research/ directory not found"

Run /orion-init <project-name> first.

"No papers found"

• Check folder path is correct
• Ensure PDFs are in the folder
• PDF files must have .pdf extension

"Can't read PDF"

Claude Code can read PDFs directly. If issues:

• Check file isn't corrupted
• Try converting to text/markdown
• Extract key sections manually

"Benchmark download fails"

• Check storage space
• Verify internet connection
• Try manual download, then /benchmark-setup --from-folder

"Hypothesis generation finds no knowledge"

Run these first:

1. /literature-review - extract from papers
2. /benchmark-setup - understand evaluation Then retry /hypothesis-generation

"Ralph Loop doesn't understand research-prd.json"

Ensure JSON format matches Ralph's expected structure:

• userStories array with hypotheses
• Each has id, title, acceptanceCriteria
• status field for tracking

Recovery Commands

Reset project:

rm -rf research/
/orion-init "project-name"

Check state:

cat research/orion.json

View all knowledge:

/knowledge status

Getting Help

1. /orion-help <topic> - specific documentation
2. /knowledge - see what's accumulated
3. Check research/orion.json for state

Tools Used

• Direct output: Display documentation to user

Note

This is a read-only command that displays inline documentation.