🤿 Deep-Ass Research

Research like a paranoid PhD, not a search bar.

Maps a topic wide, commits to depth on purpose, tries to refute its own findings, and leaves you a backlinked, primary-sourced knowledge vault — in any AI coding agent.

Before / After • Benchmarks • How it works • Install • What you get • Full install guide

---

Most "research" an agent does is one search, three skimmed snippets, and a confident paragraph that may or may not be true. DAR runs the loop a careful human would: cast wide, decide where to go deep, read the primary source (not the thread about it), send skeptics to disprove every load-bearing claim, and write it all into an Obsidian vault where every sentence traces to a source you can re-open.

It's host-agnostic: the methodology is plain markdown (core/), and thin adapters bind it to Claude Code, Cursor, opencode, Codex, or anything that reads AGENTS.md.

🪞 Before / After

You ask: "How much revenue does Perplexity make and what's their strategy?"

🔍 Normal agent "Perplexity makes around $100M and is positioned as an AI answer engine competing with Google." One search. No source. No idea if it's current, self-reported, or made up.

🤿 DAR A vault: a claim note S-001 — "~$100M ARR by mid-2025" — with the verbatim quote, the primary URL, status: verified, an independent corroboration, and a caveat that it's self-reported run-rate. Linked to [[Perplexity AI]], [[Aravind Srinivas]] (with sourced beliefs from his actual posts), and a thesis thread — plus an honest list of what's still unproven.

Same question. One is a vibe. The other is a defensible, navigable answer you can audit.

Benchmarks

How it works

DAR has two halves that share one spine — the charter (what we're actually trying to answer). A gate stops it from going deep too early; a drift guard stops it from wandering once it's deep. The charter is a compass, not a cage: useful tangents get promoted, dead ones pruned.

flowchart TD
    A(["📥 Research request"]) --> P["0 · Preflight<br/>bind capability verbs to host tools"]
    P --> C["1 · Decompose &amp; Charter<br/>success criteria SC1…SCn · scope · clarify with user"]
    C --> B["2 · Breadth<br/>parallel SEARCH-only scouts map the landscape"]
    B --> G{"Gate<br/>saturated? · rankable?<br/>anchors hit?"}
    G -->|widen one wave| B
    G -->|too broad: narrow| C
    G ==>|human check-in| D["3 · Depth loop<br/>divers FETCH full primaries<br/>→ atomic, cited claim notes"]
    D --> DG{"4.5 · Drift Guard<br/>still serving the charter?"}
    DG -->|CONTINUE| D
    DG -->|REFOCUS: prune / promote| D
    DG -->|ESCALATE: ask the user| C
    DG ==>|CONCLUDE| V["4 · Verify<br/>3 skeptics try to refute · 2-of-3 majority<br/>→ arbiter breaks ties"]
    V --> S["5 · Synthesize<br/>librarian links &amp; dedupes → synthesizer writes the map"]
    S --> O(["📦 Deliver<br/>Map-of-Content · report · open questions"])

    classDef phase fill:#0f172a,stroke:#334155,color:#e2e8f0;
    classDef gate fill:#fde68a,stroke:#b45309,color:#1c1917;
    classDef done fill:#bbf7d0,stroke:#15803d,color:#052e16;
    class P,C,B,D,V,S phase;
    class G,DG gate;
    class A,O done;

Loading

The principles behind each step (why they exist) — straight out of how good researchers actually work:

• Range before depth. Cheap, disposable SEARCH-only scouts first; expensive FETCH only on threads that survive the gate.
• Read the primary, not the summary. Divers fetch the full source (incl. appendices & limitations) into raw/; a claim sourced only to a summary is quarantined.
• Don't fool yourself. Skeptics are scored to disprove claims; disconfirming evidence is written into the note; contradictions are downgraded, never deleted.
• Synthesis is the contribution. The deliverable is a navigable _MOC.md + [[wikilink]] graph + a report answering each criterion — not a wall of text.

Install

Tip

Recommended: set up these MCP servers first, then install DAR. It runs on a host's built-in web tools, but is maximized by this free toolset — each one powers a capability verb:

MCP server	Powers in DAR
Exa	semantic / neural SEARCH — papers, people, companies
Tavily	structured web SEARCH + FETCH (date-filterable)
TinyFish	BROWSE — social / authenticated / interactive pages
Context7	READ_DOCS — version-accurate library & API docs
Ref	READ_DOCS — specs, standards, references

Set the matching keys (EXA_API_KEY, TAVILY_API_KEY, …) in your environment. On the Claude Code plugin path they're auto-wired from .mcp.json (you just enable them once). Anything missing degrades loudly, never silently — details in core/PREREQUISITES.md.

Claude Code — plugin (one command, auto-wires MCP)

/plugin marketplace add cstar0521/Deep-Ass-Research
/plugin install deep-ass-research@dar-marketplace

Registers the seven dar-* role subagents, the /deep-ass-research:dar command, and the prerequisite MCP servers.

Claude Code — skill (zero assembly)

git clone https://github.com/cstar0521/Deep-Ass-Research.git
ln -s "$(pwd)/Deep-Ass-Research" ~/.claude/skills/deep-ass-research

Then just ask for "deep research on X" or run /deep-ass-research.

Codex — plugin (one command, bundles skill + MCP)

codex plugin marketplace add cstar0521/Deep-Ass-Research

Then open Codex → /plugins → install Deep Ass Research. The DAR skill ships bundled (.codex-plugin/plugin.json + skills/), and the prerequisite MCP servers come from .mcp.json.

Cursor

cp adapters/cursor/deep-ass-research.mdc .cursor/rules/
cp adapters/cursor/dar.command.md        .cursor/commands/dar.md

opencode

cp adapters/opencode/deep-ass-research.md ~/.config/opencode/agent/
# then merge adapters/opencode/opencode.json (MCP servers + dar-* subagents)

Any other agent (Aider, Gemini CLI, custom)

Load adapters/generic/AGENTS.md, or paste adapters/generic/SYSTEM_PROMPT.md into the system prompt. No sub-agents? DAR runs the roles sequentially in-context — same method.

Full per-host details: INSTALL.md.

What you get

A self-contained Obsidian vault at ~/research/dar/<date>-<topic>-<id>/ (override with DAR_VAULT_ROOT):

In the vault	What it is
sources/S-NNN-*.md	one atomic claim per note — verbatim quote, primary URL, status + confidence, links
entities/*.md	companies / people / concepts (people carry sourced, quoted beliefs)
threads/T-*.md	synthesis notes with a stated thesis, a reasoning chain of [[links]], and counter-evidence
raw/	the cached primary sources — the provenance floor, re-openable forever
_verify-log.md	every load-bearing claim's verdict + dissent
_MOC.md + 99-report.md	the front-door map and the report answering each success criterion

Run by seven focused roles you can spawn in parallel:

Phase	Roles
Breadth	scout (cheap, wide, search-only)
Depth	deep-diver (reads full primaries)
Verify	skeptic ×3 → arbiter
Synthesize	librarian (links/dedupe) → synthesizer (the deliverable)
Steering	relevance-monitor (the drift guard)

Repo layout

deep-ass-research/
├── SKILL.md                  # Claude Code entry (drop-in skill)
├── core/                     # PORTABLE methodology — names no host tool
│   ├── capabilities.md  PREREQUISITES.md  methodology.md  orchestration.md
│   ├── vault-layout.md  note-schemas.md  provenance.md
│   ├── roles/   scout · deep-diver · skeptic · arbiter · librarian · synthesizer · relevance-monitor
│   └── modes/   academic · gtm · technical · market
├── adapters/
│   ├── generic/    AGENTS.md + SYSTEM_PROMPT.md
│   ├── claude-code/ dar-pipeline.workflow.js + plugin/   (pre-assembled, installable)
│   ├── cursor/     deep-ass-research.mdc + dar.command.md
│   ├── codex/      plugin/   (pre-assembled, installable — .codex-plugin + skills/)
│   └── opencode/   deep-ass-research.md + opencode.json + subagents/
├── .claude-plugin/marketplace.json   # makes the repo a one-command Claude Code plugin
└── .agents/plugins/marketplace.json  # makes the repo a one-command Codex plugin

Start at core/orchestration.md — the runbook every adapter points to.

Research modes

DAR picks a playbook by topic, each with its own decomposition and source strategy:

Mode	For	Move
academic	papers / literature	read methods + appendix + limitations; build a citation graph
gtm	a company, its people & beliefs	company → people → their actual posts → coherent narrative chain
technical	"how does X work", tool choices	primary docs (Context7/Ref); A-vs-B-vs-C with a recommendation
market	competitive / sector intel	players → positioning → pricing/funding/trend signals → matrix

---

_{Built to be portable, auditable, and hard to fool. Read the primary. Write everything down. Don't kid yourself.}