Deep Research

Deep Research — Systematic Research Ladder for Claude Code

> Six composable research skills (L0→L5). From a 60-second fact-check to a 40-minute academic-grade investigation with 120+ sources. Pick the right depth for the task.

[](LICENSE) [](https://claude.com/claude-code) []() [](README.ru.md)

---

What is this?

A set of six Claude Code skills that form a research ladder. Each tier composes on top of the previous one, adding specific capabilities:

| Tier | Skill | Time | Sources | Output | Best for | |:----:|-------|:----:|:-------:|--------|----------| | L0 | /quick-research | ~1 min | 3–5 | 1-paragraph answer with citations | Fact-checks, simple lookups, "latest version of X" | | L1 | /research | ~5 min | 10–15 | ~1,000-word structured report | Standard overviews, "how does X work" | | L2 | /deep-research | ~12 min | 20–30 | ~2,000-word report + contradictions | Tech choices, comparisons, non-trivial investigations | | L3 | /expert-research | ~20 min | 40–60 | ~3,000-word report + critic review | Strategic decisions, technology migrations | | L4 | /academic-research | ~40 min | 80–120 | ~5,000-word report with bibliography | Research-grade, scientific overviews | | L5 | /ultra-research | 1+ hour | 150+ | Full knowledge vault (10,000+ words) | Become an expert on a topic |

Specialized skills (parallel to ladder)

| Skill | Time | Sources | Output | Best for | |-------|:----:|:-------:|--------|----------| | /osint-research | ~10–15 min | 8–12 channels | hybrid findings/dossier/graph report | Entity recon: domain, IP, email, person, company. See docs/OSINT_INTEGRATION.md. |

Key property: L3 calls L2 which calls L1. Every higher level inherits the foundation and adds its own layer (reflection, critic, multi-perspective, academic sources, recursive loops).

---

Why does this exist?

LLM-assisted research suffers from three chronic problems:

1. Shallow by default — ad-hoc "search the web" gives you 1–2 sources and confident-sounding synthesis with no verification.
2. Hollow synthesis — the output sounds authoritative, but there is no audit trail. You can't trace a claim back to a source.
3. Context bloat — pouring research into your main conversation bloats the context window and degrades later tool use.

This ladder solves each:

| Problem | Solution | |---------|----------| | Shallow by default | Tiered depth — pick L0 for a lookup, L4 for a literature review | | Hollow synthesis | Every L1+ scrapes full content, writes per-source summaries, grades confidence | | Context bloat | All artifacts live in .firecrawl/research// — the main chat stays clean |

---

Quick Start

Prerequisites

Required:

• Claude Code CLI (any recent version)
• Firecrawl CLI — scraping engine
• Tavily MCP — search API (free tier works)

Optional add-ons (each skill degrades gracefully if absent):

• Exa MCP — neural semantic search (L2+, v0.5.0+) · free tier 1000/mo
• Perplexity MCP — answer engine (L0+ killer path, v0.6.0+)
• OpenAI Codex CLI — cross-model channel (L2+, v0.2.0+) · needs ChatGPT Pro

Full setup: see [docs/INSTALLATION.md](docs/INSTALLATION.md).

Install

Option 1 — Claude Code plugin (v0.4.0+, one-liner):

/plugin install hint-shu/deep-research

Installs all six skills, the shared verification library, and the Codex helper in one command. Requires Claude Code 2.1.30+.

Option 2 — manual (works on any Claude Code version):

git clone https://github.com/hint-shu/deep-research.git
cd deep-research
bash scripts/install.sh

The script copies skills/* into ~/.claude/skills/, deploys the shared lib to ~/.claude/scripts/lib/, and prints next-step instructions for API keys.

First run

/quick-research latest stable version of Bun
/research how does Tailwind CSS v4 differ from v3
/deep-research PostgreSQL vs MongoDB for an e-commerce platform in 2026

That's it. Each skill writes its artifacts to .firecrawl/research// in whatever directory Claude Code is currently running in.

---

How it works

Each skill is a markdown file that Claude Code loads as instructions. When you invoke a skill, Claude follows its pipeline:

L3 Expert Research
  └─ calls L2 Deep Research
       └─ calls L1 Base Research
            ├─ 1. Plan     — decompose query into 3 sub-questions
            ├─ 2. Search   — parallel Firecrawl + Tavily for each sub-q
            ├─ 3. Scrape   — full page content, one file per URL
            ├─ 4. Summarize— per-source .md summary (title, findings, citations)
            └─ 5. Report   — ~1,000-word synthesis with bibliography
       ├─ 6. Reflection    — find gaps in L1 report
       ├─ 7. Follow-up     — 10–15 new sources for gaps
       ├─ 8. Contradiction — surface disagreements between sources
       └─ 9. Confidence    — grade each claim H/M/L
  ├─ 10. Critic agent      — independent pass, challenges conclusions
  ├─ 11. Neutral angle     — fresh perspective search
  └─ 12. Executive summary — 3,000-word final report

Artifact layout after an L3 run:

.firecrawl/research/postgres-vs-mongo-2026/
├── L1/
│   ├── plan.md              ← sub-questions
│   ├── search-1.json        ← raw search results
│   ├── sources/             ← full scrapes
│   │   ├── 01-aws-docs.md
│   │   └── 02-mongo-atlas.md
│   ├── summaries/           ← per-source digests
│   └── report.md            ← L1 synthesis
├── L2/
│   ├── gaps.md              ← reflection output
│   ├── sources/             ← new scrapes
│   ├── contradictions.md
│   └── report.md            ← L2 synthesis
└── L3/
    ├── critic-review.md     ← adversarial pass
    ├── neutral-angle.md
    └── report.md            ← final executive report

---

Real use cases

Software engineering

• Technology choices — /deep-research Next.js vs SvelteKit for our e-commerce startup
• Library migration — /deep-research migrating from Prisma to Drizzle — breaking changes and gotchas
• New release research — /research what's new in React 19 and production readiness
• Architecture patterns — /expert-research should we use CQRS for our multi-tenant SaaS

Business & strategy

• Market analysis — /expert-research state of Telegram Ads for gaming apps in 2026
• Competitive intel — /deep-research how does Linear compare to Jira for 20-person engineering team
• Pricing research — /research pricing models for B2B AI SaaS in 2025-2026
• Regulatory — /academic-research EU AI Act compliance for LLM-based products

Academic & journalism

• Literature review — /academic-research MoE architectures in LLMs 2023-2026
• Fact-checking with trail — /deep-research claims about GPT-5.4 BrowseComp 89.3% benchmark
• Trend analysis — /expert-research rise and decline of crypto-based social networks
• Multi-perspective — /academic-research remote work productivity studies methodologies

Full library with example prompts and expected outputs: see [docs/USECASES.md](docs/USECASES.md).

---

What's special about this

| Feature | What it gives you | |---------|-------------------| | Composable ladder | Each tier inherits the full pipeline of tiers below. No copy-pasted logic. | | Cross-model verification (L2+) | Optional Codex CLI channel adds GPT-5.4 as a second, independent research voice. Different search index, different model, fewer blind spots. | | Confidence grading | Every L2+ claim is tagged H/M/L. You know which parts of the synthesis to trust. | | Contradiction surfacing | L2 explicitly writes a contradictions.md — no more hidden disagreements between sources. | | Critic agent (L3+) | Separate agent reviews the report adversarially before you see it. | | Per-source summaries | Never lose the audit trail — every cited fact points to a full scrape on disk. |

---

Documentation

| Doc | What's inside | |-----|---------------| | [docs/INSTALLATION.md](docs/INSTALLATION.md) | Full setup on macOS/Linux/WSL, all API keys, verification | | [docs/USAGE.md](docs/USAGE.md) | Every skill explained with inputs, outputs, and tips | | [docs/USECASES.md](docs/USECASES.md) | 20+ real-world scenarios with ready-to-copy prompts | | [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) | Design philosophy — why a ladder, why these boundaries | | [docs/CODEXINTEGRATION.md](docs/CODEXINTEGRATION.md) | Optional GPT-5.4 cross-model channel for L2+ | | [docs/TROUBLESHOOTING.md](docs/TROUBLESHOOTING.md) | Common issues and fixes | | [CONTRIBUTING.md](CONTRIBUTING.md) | How to add skills, improve pipelines, open PRs |

---

Requirements at a glance

| Component | Required? | Free tier? | Notes | |-----------|-----------|------------|-------| | Claude Code | Yes | — | Core runtime | | Firecrawl | Yes | 500 scrapes/mo | For all levels | | Tavily | L1+ | 1,000 searches/mo | Skip only if you use L0 exclusively | | Codex CLI | No | Requires ChatGPT Pro | Unlocks L2+ cross-model channel |

Total cost if you stick to free tiers: $0. With Codex: existing ChatGPT Pro ($200/mo) — no extra API fees.

---

Roadmap

Shipped:

• [x] v0.1 — Six-tier research ladder (L0→L5), Firecrawl + Tavily
• [x] v0.1.1 — L2 verification checkpoints (anti-"hollow synthesis")
• [x] v0.2 — Optional Codex CLI cross-model channel
• [x] v0.3 — Shared verification library + flagship live example
• [x] v0.4 — Claude Code plugin manifest + skill migration
• [x] v0.5 — Exa MCP (neural semantic search)
• [x] v0.6 — Perplexity MCP (answer engine — L0 killer path)

Planned:

• [ ] v0.7 — Kagi as 5th backend for consumer research (anti-SEO-spam)
• [ ] v0.7 — Streaming Codex output (reduce L3 latency)
• [ ] v0.7 — MCP-based Codex integration (replace Bash shell-out)
• [ ] v0.8 — Full Russian translation of ARCHITECTURE + TROUBLESHOOTING
• [ ] v1.0 — Obsidian auto-sync for L5 vaults

See [CHANGELOG.md](CHANGELOG.md) for what's shipped.

---

Contributing

Contributions are welcome — new skills, new backends, better pipelines, translations. Start with [CONTRIBUTING.md](CONTRIBUTING.md).

Good first issues:

• Improve L2 verification checkpoints (see issue tracker)
• Add a new search backend (Kagi, Exa, Brave Search)
• Write use-case examples for a domain we haven't covered
• Translate docs to your language

---

License

[MIT](LICENSE) — use freely, modify freely, attribution appreciated but not required.

---

Credits

Built on top of:

• Claude Code — Anthropic's CLI
• Firecrawl — web scraping engine
• Tavily — research-grade search API
• OpenAI Codex CLI — optional cross-model channel

Originally designed by @hint-shu as an internal research toolkit, open-sourced because good systematic research should not be a closely-guarded secret.

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: hint-shu
• Source: hint-shu/deep-research
• License: MIT

Install and usage instructions live in the source repository linked above.