Deep Research

Deep Research

Overview

deep-research runs a recursive, multi-round web investigation. It fans out parallel subagents breadth-first then depth-first, runs an adversarial verification valve once per round, and synthesizes by re-deriving contested claims from retained raw sources rather than from a summary chain. The loop is a double-diamond: each round alternates diverge (explore/expand) with converge (synthesize/review), repeating until findings saturate or the user's budget runs out.

Four invariants make this safe and cheap — they are load-bearing, not optional:

1. Raw-to-disk. Every fetch is dumped verbatim to a tmpfile. Subagents return only {claim, source-url, raw-pointer, self-tag} — never raw content upward.
2. Review re-derives from raw. The verification valve checks contested/load-bearing claims against the raw tmpfiles, never against the distilled findings ledger.
3. Review fans out by claim. "Once per round" is a cadence, not a single monolithic pass: extract atomic claims, dispatch one skeptic per claim (or small batch).
4. Synthesis reads artifacts; the user reads progress. Progress prose is for the human and must never become a synthesis input — the ledger + topic-map + raw are the lossless memory.

Harness Assumptions

This skill assumes three harness primitives. Substitute your harness's equivalents and gate each on availability:

• Parallel subagent dispatch — the fan-out engine for explore/exploit/verify. If unavailable, run dimensions sequentially in-context, but still honor raw-to-disk so the synthesis window stays clean. This skill's dispatch cadence, model-tier map, and verification valve are purpose-built for research and take precedence over any general subagent skill (e.g. subagent-patterns) during a research run.
• Web search and fetch tools — native or MCP (see Tool Selection).
• A user-question primitive for Round 0 (for example AskUserQuestion). If unavailable, ask in plain text and wait.

When to Use

• A question that needs many sources, cross-checking, and a structured cited report.
• "What's the current state / landscape / consensus on X?"
• Claims that must be verified, not just collected; topics with active disagreement.
• Time-sensitive topics where recency and temporal precision matter.

When Not to Use

• Single-fact lookups or quick answers — just search directly.
• Library, API, SDK, dependency, or version questions — use mcp-research.
• Open-ended ideation with no anchor — use brainstorming.
• Reading/auditing a local codebase — use code search tools.

Invocation Notice

• Tell the user when this skill is running: deep-research. Skip the notice when the user asked for the skill by name or slash command; spelling and spacing need not match. A phrase from this skill's own trigger list is not a name — naming the work is not naming the skill.

The Loop

digraph deep_research_loop {
    rankdir=TB;
    "Research request" [shape=doublecircle, style=filled, fillcolor=lightgreen];
    "Question specific enough?" [shape=diamond];
    "R0: Ask <=5 clarifying questions (single volley)" [shape=box];
    "R1: Frame - date, decompose, topic-map, set caps" [shape=box];
    "R2: Explore breadth (parallel subagents) -> dump raw" [shape=box];
    "R3: Exploit depth (parallel subagents) -> dump raw" [shape=box];
    "Review valve: verify claims vs RAW, fan out by claim" [shape=diamond];
    "R4: Expand/Resolve - targeted re-fetch + re-review" [shape=box];
    "Report progress to user" [shape=box];
    "Loop cap reached or saturated?" [shape=diamond];
    "NEVER inline raw into synthesis" [shape=octagon, style=filled, fillcolor=red, fontcolor=white];
    "Synthesize report from artifacts + raw" [shape=box];
    "Cited report delivered" [shape=doublecircle, style=filled, fillcolor=lightgreen];

    "Research request" -> "Question specific enough?";
    "Question specific enough?" -> "R1: Frame - date, decompose, topic-map, set caps" [label="yes"];
    "Question specific enough?" -> "R0: Ask <=5 clarifying questions (single volley)" [label="no"];
    "R0: Ask <=5 clarifying questions (single volley)" -> "R1: Frame - date, decompose, topic-map, set caps";
    "R1: Frame - date, decompose, topic-map, set caps" -> "R2: Explore breadth (parallel subagents) -> dump raw";
    "R2: Explore breadth (parallel subagents) -> dump raw" -> "R3: Exploit depth (parallel subagents) -> dump raw";
    "R3: Exploit depth (parallel subagents) -> dump raw" -> "Review valve: verify claims vs RAW, fan out by claim";
    "Review valve: verify claims vs RAW, fan out by claim" -> "Synthesize report from artifacts + raw" [label="claims hold"];
    "Review valve: verify claims vs RAW, fan out by claim" -> "R4: Expand/Resolve - targeted re-fetch + re-review" [label="gaps / contested"];
    "R4: Expand/Resolve - targeted re-fetch + re-review" -> "Report progress to user";
    "Report progress to user" -> "Loop cap reached or saturated?";
    "Loop cap reached or saturated?" -> "R2: Explore breadth (parallel subagents) -> dump raw" [label="no + coverage gap"];
    "Loop cap reached or saturated?" -> "Synthesize report from artifacts + raw" [label="yes / saturated"];
    "NEVER inline raw into synthesis" -> "Synthesize report from artifacts + raw" [style=dotted, label="invariant"];
    "Synthesize report from artifacts + raw" -> "Cited report delivered";
}

Rendered: assets/the-loop.svg.

Per-round mechanics, artifact schemas, dispatch templates, and the model-tier map: see references/round-playbook.md.

The Review Valve

Fires once per round at the converge moment (end of R3, and on R4 material) — not per subagent. It is the seam where the loop (R3/R4) reads the memory (raw tmpfiles) under the rule (re-derive, don't trust the summary). Verify only load-bearing, contested, surprising, or single-source claims — not everything. Each verified claim gets a verdict (Supported / Partial / Unsupported / Uncertain) and a confidence. Sufficient unresolved questions trigger R4. Full procedure, claim-selection criteria, verdict rubric, and conflict-resolution rules: see references/verification.md.

Tool Selection — use everything available

Inventory the search/retrieve tools present at runtime and prefer the richest set — do not default to whatever the harness ships natively. A dedicated research or extraction tool (often MCP-provided) usually beats a generic web fetch on coverage and clean output. If tools are deferred or hidden, discover and load them first (for example via a tool-search primitive) before dispatching subagents.

Map the tools you have to these capability needs; native search/fetch is the floor when nothing richer is available:

Directives:

• Read full sources, not snippets. Search snippets locate; only a full fetch grounds a claim.
• Cross-source corroboration is mandatory for web claims — no native provenance to lean on.
• Subagents own all token-heavy retrieval; the orchestrator and synthesizer never call retrieval tools directly (keeps the synthesis window clean — invariant 1).

Keeping the User Updated

• After Round 1, state the plan in one short message: dimensions, loop cap, rough scope.
• Emit a progress checkpoint after each R4 loop: a diff of the ledger — what resolved this loop, what is still contested, "loop N of cap M", and an explicit continue-or-synthesize prompt. This is the human-in-the-loop stop gate and where the budget cap is honored interactively.
• Keep checkpoints terse and derived from the ledger. Never re-summarize source content into them, and never feed checkpoint prose back into synthesis (invariant 4).

Output

A cited report (see references/temporal-and-output.md for the full contract):

• Organized around themes, not walked through source by source.
• Graded epistemics: established / active debate / speculation, with confidence levels.
• Conflicts presented side-by-side with verdicts — never silently collapsed to one answer.
• Every claim carries a pointer to its source URL (and raw tmpfile during the run).
• Explicit gaps and limitations section.

Related Skills

Hand off or compose at these flow points:

• brainstorming — before R0, when the question itself is unformed and you do not yet know what to research. R0 only narrows a researchable question; it does not shape a fuzzy premise.
• mcp-research — mid-research, when a sub-question is a library, API, SDK, or version lookup. Its docs-first tool selection applies within that thread.
• antislop-writing — at Synthesize, to tighten the report. The report is prose a human reads, so its heading-claim and lead-with-the-point rules apply (they do not apply to this skill file).
• editorial-review — after the report, only if it advances a thesis or recommendation worth pressure-testing. A pure findings catalog does not need it.

References

• references/round-playbook.md — per-round steps, artifact schemas (topic-map, findings-ledger, tmpfile/pointer), subagent dispatch templates, model-tier map, saturation definition.
• references/verification.md — review valve procedure, claim selection, verdict taxonomy, evidence-strength, conflict resolution, single-source/replication flags.
• references/temporal-and-output.md — intent→temporal-precision table, date phrasing, report contract and skeleton.