Autonomous research agent skill for Claude Code. Give it a topic, and it autonomously searches the web, verifies sources, synthesizes findings, and produces a structured research report — all running in the background while you continue working.
/omni-research
Asking an AI to "research X" gives you a one-shot answer based on training data. Omni-Research is different: it runs a multi-cycle autonomous loop that searches the web, reads actual pages, verifies every claim at the source, and builds a growing knowledge base over 5-30+ research cycles. The result is closer to what a human research analyst would produce — with citations you can actually check.
You Claude Code Background Agent
| | |
| /omni-research | |
|------------------------------>| |
| "Research X topic" | |
| | Interactive setup: |
| <-- Topic? Type? Scope? --> | 1. Refine topic |
| <-- Research lines? -------> | 2. Classify type |
| <-- Confirm? --------------> | 3. Propose research lines |
| | 4. Configure depth |
| | |
| | Launch background agent --------->|
| | |
| "Agent launched. Check | PHASE 0: Source Scouting |
| research.md anytime." | - Hunt curated lists, blogs, |
| | community threads |
| (continue your work) | - Score sources (1-5) |
| | - Build Source Queue (8+ items) |
| | |
| (edit steer.md to redirect) | PHASE 1: Seed Mining |
| | | - WebFetch & deep-read sources |
| v | - Extract findings with inline |
| "Focus on Line 3" | citations |
| | |
| | PHASE 2: Hypothesis Loop |
| | 1. Check steer.md for redirects |
| | 2. Pick biggest knowledge gap |
| | 3. WebSearch for evidence |
| | 4. WebFetch & verify sources |
| | 5. Synthesize with inline cites |
| | 6. Update research.md |
| | 7. Every 4th: PIVOT/REFINE gate|
| | 8. Every 5th: discover sources |
| | 9. Repeat until saturated |
| | |
| | COMPLETION: |
| | - Anti-fabrication verify pass |
| <-- "Research complete!" ----|--- Generate BRIEF.md --------------|
| | |
- 4 research types — product, marketing, production, pure research — each with tailored BRIEF templates
- Autonomous loop — searches, reads, evaluates, and synthesizes without intervention
- Source verification — every claim is WebFetched and verified in the original page before citing (no search-snippet-only citations)
- Structured output —
research.md(growing knowledge base),experiments.tsv(cycle log),BRIEF.md(executive summary) - Configurable depth — quick survey (5-8 cycles), comprehensive (12-20), or deep dive (20-30+)
- Mid-run steering — edit
steer.mdanytime to redirect the agent: change focus, add context, skip lines, or wrap up early - Inline citations — every claim links to its source inline, not just in an appendix
- Confidence signals — findings marked with consensus (✓ agree / ~ mixed / ✗ conflict) and confidence levels (High/Medium/Low)
- Source scoring — sources rated on authority, recency, and relevance (1-5 each) to prioritize quality
- PIVOT/REFINE gates — every 4 cycles the agent evaluates progress and can change direction instead of grinding on dead ends
- Anti-fabrication pass — final sweep verifies every citation URL actually appeared in search results before publishing
- Auto-detected language — writes output in the same language you use
- Live monitoring — check
research.mdanytime to see progress - Recovery mode —
/omni-research brief <path>regenerates BRIEF.md from existing research data
Install as a Claude Code plugin for auto-updates and easy management. This repo is a self-hosted marketplace — add it once, then install the plugin:
/plugin marketplace add romanticamaj/omni-research
/plugin install omni-research@omni-research-marketplace
Then restart Claude Code. Verify with /plugin list.
macOS / Linux:
git clone https://github.com/romanticamaj/omni-research.git
cp -r omni-research ~/.claude/skills/omni-researchWindows:
git clone https://github.com/romanticamaj/omni-research.git
xcopy /E /I omni-research %USERPROFILE%\.claude\skills\omni-researchStart a new Claude Code session and type:
/omni-research
The first time you run it, it will ask where to save research output and remember your choice.
/omni-research
Follow the interactive prompts:
- Topic — describe what you want to research
- Type — product / marketing / production / research (auto-detected)
- Context — share intuitions, links, files, or reference materials
- Research lines — review and customize proposed research directions
- Scope — quick (a), comprehensive (b), or deep dive (c)
- Confirm — the agent launches in background
While the agent runs, check these files in your output directory:
| File | Description |
|---|---|
research.md |
Growing knowledge base — updated every cycle |
experiments.tsv |
One row per cycle: hypothesis, sources found, key finding, status |
steer.md |
Edit this to redirect the agent mid-run (change focus, add context, wrap up) |
program.md |
The full research program (for reference) |
BRIEF.md |
Executive summary (generated when research completes) |
Edit steer.md in the output directory to redirect the agent while it's running:
Focus on Line 3 — I found this article that might help: https://example.com/articleThe agent checks this file at the start of every cycle and follows your instructions.
If the session ended early or you want to regenerate the summary:
/omni-research brief path/to/research-output/
<output_base_dir>/
2026-03-28-suno-prompting-r7k2/
program.md # Research instructions
research.md # Accumulated knowledge (grows each cycle)
steer.md # Steering file (edit to redirect agent)
experiments.tsv # Experiment log
BRIEF.md # Executive summary (generated at completion)
omni-research/
.claude-plugin/
plugin.json # Plugin manifest (name, version, config schema)
skills/
omni-research/
SKILL.md # Skill definition (interactive flow)
program-template.md # Template for the autonomous agent's instructions
config.json.example # Example config
templates/
brief-product.md # BRIEF template for product research
brief-marketing.md # BRIEF template for marketing research
brief-production.md # BRIEF template for production research
brief-research.md # BRIEF template for pure research
README.md
LICENSE
- Claude Code CLI, desktop app, or IDE extension
- A Claude plan with background agents and web tools (WebSearch, WebFetch) — this includes Max and Team plans
| Issue | Cause | Fix |
|---|---|---|
| Agent stops mid-research | Context window limit reached | Use /omni-research brief <path> to generate BRIEF from partial results |
| "WebSearch not available" | Plan doesn't support web tools | Upgrade to a plan with web tool access |
| Agent drifts from topic | Long research sessions can lose focus | Edit steer.md to redirect, or reduce scope to "quick survey" |
config.json not found |
First-time setup | The skill will prompt you — just provide the output directory path |
Empty research.md after launch |
Agent failed to start | Check that program.md was created in the output directory |
MIT