| name | deep-research |
| description | Use when the user asks you to research a topic and wants thesis-grade synthesis, not a bullet-point summary. Activates on "research X", "deep dive on X", "find out about X", "what's actually known about X". Dispatches depth + breadth + counter-argument subagents concurrently, runs a peer-review pass over their findings for hallucination and citation quality, then synthesizes into a structured report with explicit high-confidence / contested / unknown buckets. Output file optionally persisted to a knowledge vault. |
Deep Research
A 4-role, peer-reviewed research pattern that produces thesis-grade output
instead of the web-search summary slop that LLMs default to.
When to activate
- User asks "research X", "deep dive on X", "find out about X"
- User says "I want to know what's actually known about X"
- User asks a question where an answer built from one web search or
confident-guessing from training data would be dangerous (load-bearing
decisions, claims that will be cited, topics where hallucinations are
likely)
- A previous conversation surfaced a speculative claim that needs
validation before it lands in code or a plan
When NOT to activate
- Quick factual lookups ("what's the latest version of React") — use
Context7 or a single web fetch
- Well-known concepts the model already knows cold — answer directly
- Questions about the local codebase — use the Explore agent
- The user wants opinion, not evidence
The four roles
All four are dispatched via the Agent tool with subagent_type: general-purpose.
Each role has a distinct prompt shape — do not collapse them.
1. Depth researcher (1 agent)
Goes deep on the PRIMARY question. Multiple web searches + web fetches
on the specific topic. Produces a detailed findings dossier with citations
for every load-bearing claim.
2. Breadth researcher (1 agent)
Surveys adjacent/related topics, alternative framings, and historical
context. Broader search scope, looking for relevance-adjacent material
the depth agent might miss. Explicitly asks "what related things should
I know to understand this properly?"
3. Counter-argument researcher (1 agent)
Actively searches for contradicting views, cases where the thesis fails,
opposing schools of thought, known criticisms. Not "balance" — actively
adversarial, looking for holes.
4. Peer reviewer (1 agent, runs AFTER the first three return)
Takes all findings from roles 1-3 and produces a peer-review report:
- Claim → citation mapping (which claims have support, which are bare assertions)
- Internal contradictions between findings
- Claims where multiple sources agree (high-confidence)
- Claims where sources disagree (contested)
- Gaps — things that were researched but no citation was found
- Hallucination risk flags — claims that sound specific but have no citation
Process
- Brief. Read the user's topic. Extract:
- The primary question (depth target)
- 2-4 adjacent topics (breadth targets)
- 1-2 counter-hypotheses to actively seek (counter targets)
- Any specific source types to privilege (academic papers, primary
sources, industry reports) or exclude (pure marketing content)
- Dispatch concurrently. In ONE message, send three Agent tool calls
(depth + breadth + counter). They run in parallel. Each returns a
structured findings dossier.
- Dispatch peer reviewer with all three dossiers as input.
- Synthesize (main session, not subagent). Produce the report per
the output format below.
- Persist to the vault if one exists (see "Vault persistence").
See references/subagent-briefs.md for the exact prompt template each
role should receive.
Output format
Every deep-research output has the same six-section shape. See
references/output-format.md for the full template.
- Bottom-line-up-front — 2-3 sentences answering the question with
an explicit confidence grade.
- High-confidence findings — claims with multi-source corroboration.
- Contested claims — where sources disagree; both sides with
citations.
- Unknowns — questions the research did NOT resolve.
- Follow-up research questions — what to investigate next.
- Citations — consolidated list, graded by source quality.
Hallucination flags and peer-review notes appear inline next to the
specific claim, not in a separate dumping ground. A report that reads
clean without any flags is a report the peer-reviewer didn't try hard
enough on.
Vault persistence
If the user has an Obsidian-style vault (default path:
~/Documents/Personal Notes/), write the report to
50-References/deep-research-{slug}-{YYYY-MM-DD}.md and update
index.md + log.md per the standard vault-write convention.
If no vault is configured, write to ./research/{slug}-{YYYY-MM-DD}.md
in the current working directory.
In both cases, report the output path back to the user so they can
find it later.
Concurrency rules
- Depth + breadth + counter dispatch IN PARALLEL (single message, 3 tool
calls). Do not sequence them.
- Peer reviewer runs AFTER, sequentially. It needs the first three's
outputs as input.
- Maximum 3 concurrent subagents at the research stage. Do not expand.
- Max 1 deep-research invocation at a time per session — if the user
asks for two topics, do them sequentially.
Known failure modes
See references/failure-modes.md for details. Headline list:
- LLM-marketing-content swamp: web searches on AI/tech topics often
return SEO-stuffed blog content. Peer-reviewer must flag claims
supported only by such sources as low-grade.
- Hallucinated paper titles: subagents sometimes cite plausible-
sounding papers that don't exist. Peer reviewer verifies at least the
top 3 citations by attempting to resolve them.
- Stale research: on fast-moving topics, research older than 12
months may be obsolete. Peer reviewer calls out date freshness.
- False balance: equal weight for mainstream + fringe views is
wrong when the evidence is asymmetric. Peer reviewer judges weight.
- Scope creep: subagents drift from the primary question. The brief
given to each role must pin scope hard.
What to tell the user on completion
- One-paragraph executive summary (copy from the report's BLUF section)
- Path to the full report
- Any peer-review-flagged items that specifically bear on what they
asked (don't swallow contested claims silently)
- A recommended next step based on what the research surfaced
Do not paste the full report inline unless the user asks. The vault
(or local file) is the artifact.