Deep Research (knowledge-graph-adapted)

Adapted from 199-biotechnologies/claude-deep-research-skill (upstream Boris Djordjevic, MIT, ~633★). Verdict on upstream: PARK + adapt. the methodology is sound but its install pattern (git clone main + cp -r) violates the security-first no-unpinned-bootstrap rule (no unpinned bootstrap, no main-branch clones). This file is the in-tree adaptation. no clone, no external script fetching.

Distinguishes from a single-pass web-search skill: a single-pass search skill is for quick "what's the latest on X" questions. This skill is multi-phase with critique/refine cycles and persistent evidence trails. If the question takes <5min and a single search settles it, route to the single-pass skill. If it takes >10min and needs synthesis across sources, route here.

Decision tree

Request analysis
├── Simple lookup?              → STOP: route to a single-pass web-search skill or WebSearch
├── Debugging?                  → STOP: use standard tools
└── Complex multi-source needed → CONTINUE

Mode selection
├── Initial exploration         → quick    (3 phases, 2-5 min)
├── Standard research           → standard (6 phases, 5-10 min)  [DEFAULT]
├── Critical decision           → deep     (8 phases, 10-20 min)
└── Comprehensive review        → ultradeep (8+ phases, 20-45 min)

If it's unclear which mode, default to standard. Surface the mode pick to the operator in 1 line before starting (e.g. "running standard, will upshift to deep if scope expands").

Phase pipeline

Phases 3-5 are an evidence loop per sub-question, NOT strict sequential gates. If a finding falls apart at TRIANGULATE, drop back to RETRIEVE.

Phase details

Phase 1: SCOPE

Output a 4-line scope block before any retrieval:

QUESTION: <restated in your own words>
ASSUMPTIONS: <2-4 bullets. what's taken as given, flag any that materially change the answer if false>
OUT-OF-SCOPE: <1-3 bullets. what we're NOT answering>
SUCCESS CRITERIA: <what makes the report useful>

If high-materiality assumptions exist (e.g. "assuming a specific regulatory interpretation as of a 2024 ruling still holds"), surface them in the final report's Introduction. not buried.

Phase 2: PLAN

Decompose into 4-8 sub-questions. Each sub-question gets:

• Source classes expected (regulatory document / academic paper / on-chain data / news / industry report / first-party project doc)
• Cluster independence target. how many cluster-independent sources before a claim is trustable. Default 3.
• Verification primitive. what makes a claim verifiable for THIS sub-question (a regulation-number lookup / an on-chain query / cross-source date triangulation / first-party confirmation)

Phase 3: RETRIEVE

Source order of preference:

1. A web-search MCP (e.g. mcp__exa__web_search_exa, mcp__exa__crawling_exa). primary. Cheap per-query cost.
2. A browser-automation MCP (for JS-rendered, auth-gated, or structured-extraction targets. region-specific dashboards or regulator pages that block a plain crawler).
3. WebFetch. fallback for known-good static pages.
4. The local knowledge graph. recall against the recall index + direct read of memory/Context/ / knowledge/ / your graph export. Always run BEFORE going external (brain-first per CLAUDE.md).
5. Domain MCPs (academic search, docs, clinical trials, etc.). if domain-relevant.

For each retrieval, capture:

• URL or local path
• Fetch timestamp (publication date if available)
• Source type (web / academic / regulatory / on-chain / industry-report / first-party)
• Authors (if attributable)
• Credibility prior (0.0-1.0). based on source class and provenance
• Verification status: unverified (just fetched) → verified (cluster-independent confirmation) → conflicted (sources disagree, flag for Phase 4)

Phase 4: TRIANGULATE

For every load-bearing claim:

• Are there ≥3 cluster-independent sources?
• Are dates consistent across sources?
• Do regulatory citations resolve in official registries? On-chain claims resolve on a block explorer (e.g. Etherscan) / analytics site (e.g. DeFiLlama)?
• If a conflict is found → flag it, surface it in the Limitations & Caveats section, do NOT silently pick a side.

Phase 4.5: OUTLINE REFINEMENT

Re-organize before SYNTHESIZE based on what evidence ACTUALLY exists. The original PLAN was a hypothesis; the outline now reflects the evidence shape. If a sub-question turned out unsupportable, demote it to Limitations.

Phase 5: SYNTHESIZE

Write findings with inline citations. Format per finding:

- finding-claim:: <one-sentence claim>
- confidence:: high | medium | low
- sources:: [[Source/<id>]] [[Source/<id>]] [[Source/<id>]]
- ## Evidence
  - <2-4 paragraphs of evidence with inline citation refs>

Quality bar:

• ≥10 sources for the full report
• ≥3 cluster-independent per major claim
• All factual claims cited immediately
• Prose-first (≥80% prose, not bullet-spam)
• No fabricated citations. verify every URL still resolves before locking the report

Phase 5.5: SOURCE-GROUNDING GATE

Lifted from a deep-research checklist (affaan-m/everything-claude-code). The quality bar above is a target; this is the hard pass/fail before anything leaves the skill. Run it as an explicit gate, not a vibe check.

Per-claim sweep on the assembled findings:

1. Every claim sourced. no asserted-as-fact claim without a sources:: ref. An unsourced claim = delete or demote to an explicit (unverified inference).
2. Single-source flag. any claim resting on exactly one source gets tagged single_source:: true. Collect all such claims into a single_source_claims:: list in the report frontmatter. This list is the explicit handoff to Phase 6 / the Critic (a pre-ship catch, not a post-hoc observation).
3. Fact vs estimate split. numbers that are measured/reported stay as-is. Numbers that are extrapolated, modeled, or "directionally" derived get an explicit (estimate) marker inline. Never present an estimate with the typographic confidence of a measured figure.
4. Recency-first. prefer sources <12mo. Any load-bearing claim resting on a >12mo source gets date_risk:: <source-date>.
5. Gap acknowledgment. if a sub-question could not be grounded, write it into a ## Open / Ungrounded section. Do not paper over it with a plausible-sounding unsourced sentence.

The gate fails (block packaging, loop back to RETRIEVE) if: any unsourced asserted-fact remains, OR single_source_claims:: covers a load-bearing thesis claim with no acknowledgment. The gate passes with single_source_claims:: populated as long as they are flagged, not hidden.

Phase 6: CRITIQUE (deep + ultradeep only)

Adversarial pass. Spawn a subagent with this brief (or do inline if <5min):

• Steel-man the OPPOSITE thesis. What would a well-informed skeptic say?
• Weakest claim. which finding has the thinnest evidence?
• Sourcing gaps. where are we relying on a single source?
• Date-stale claims. anything that may have changed in the last 6mo?
• Cluster dependence. sources that look independent but trace to the same primary?

Output: a 5-15 bullet critique. Save as outputs/raw/agent/research-critique/[topic-slug]-critique.md if substantive.

Phase 7: REFINE (deep + ultradeep only)

Fold the critique into v2:

• Strengthen claims where the critique was right
• Add Limitations bullets where the critique was right but research depth doesn't allow a fix
• Push back where the critique was off-base (note the rebuttal in the report)

Phase 8: PACKAGE (knowledge-graph-native output)

Main page path: pages/Research___YYYY-MM-DD___[topic-slug].md (relative to your graph root).

Use the graph's property block format (NOT YAML frontmatter):

- type:: [[Research]]
- date-created:: [[YYYY-MM-DD]]
- date-modified:: [[YYYY-MM-DD]]
- mode:: quick | standard | deep | ultradeep
- status:: draft | complete
- source-count:: N
- tags:: #research #[topic-tag] #[topic-tag-2]
- ## Executive Summary
  - <200-400 words>
- ## Introduction
  - ### Scope
    - <restated question>
  - ### Methodology
    - <mode + phase summary>
  - ### Assumptions
    - <high-materiality assumptions>
- ## Findings
  - ### [[Finding/<title-1>]]
    - <inline summary, full content on the linked page>
  - ### [[Finding/<title-2>]]
    - ...
- ## Synthesis
  - <cross-finding insights>
- ## Limitations & Caveats
  - <what we couldn't confirm, conflicting sources, date-stale risks>
- ## Recommendations
  - <if action-oriented>
- ## Bibliography
  - [[Source/<id-1>]]
  - [[Source/<id-2>]]
- ## Methodology Appendix
  - <mode used, sub-questions, source classes, cluster-independence achieved>

Per-finding pages at pages/Finding___YYYY-MM-DD___[finding-slug].md:

- type:: [[Finding]]
- date-created:: [[YYYY-MM-DD]]
- confidence:: high | medium | low
- parent-research:: [[Research/YYYY-MM-DD/topic-slug]]
- sources:: [[Source/<id>]] [[Source/<id>]] [[Source/<id>]]
- tags:: #finding #[topic-tag]
- ## Claim
  - <one-sentence claim>
- ## Evidence
  - <full evidence paragraphs with inline citations>
- ## Counter-evidence
  - <if any, even brief>

Per-source pages at pages/Source___YYYY-MM-DD___[source-slug].md:

- type:: [[Source]]
- url:: <https://...>
- source-type:: web | academic | regulatory | on-chain | industry-report | first-party
- date-published:: [[YYYY-MM-DD]] OR undated
- date-accessed:: [[YYYY-MM-DD]]
- authors:: <names if attributable>
- credibility:: 0.0-1.0
- verification-status:: unverified | verified | conflicted
- tags:: #source #[topic-tag]
- ## Summary
  - <2-3 sentences>
- ## Quoted excerpts
  - <exact quotes used in findings, verbatim>

Per-entity stubs at pages/Entity___[Name].md (auto-create on first reference):

- type:: [[Entity]]
- date-created:: [[YYYY-MM-DD]]
- aliases::
- source-research:: [[Research/YYYY-MM-DD/topic-slug]]
- tags:: #entity #auto-created #needs-review
- ## Summary
  - <2-3 sentences from research context>

Cross-link rules:

1. Every organization, project, technology, person, or concept in findings → [[Entity/<Name>]] syntax. If the page doesn't exist, create a stub.
2. Every finding → linked from the main report's Findings section.
3. Every source → linked from the main report's Bibliography.
4. If a finding references a concept covered in a previous research run → link to that finding page directly.

Quality gates (run before status:: complete)

• ≥10 sources, ≥3 cluster-independent per major claim
• All factual claims cited immediately
• No placeholders, no fabricated URLs
• Prose-first (≥80%)
• All entity references wiki-linked
• The Bibliography lists EVERY source actually cited
• The Limitations section honestly enumerates what we couldn't confirm
• Date-published captured for every source (or "undated" explicit)
• Regulatory citations cross-checked in official registries
• On-chain claims cross-checked on-chain (a block explorer / analytics site)

If any gate fails → status remains draft, surface the failures in a - ## Open issues block at the bottom of the main page.

When to use / NOT use

Use:

• Comprehensive analysis spanning multiple sources (>5 sources expected)
• Technology comparisons requiring structured findings
• Regulatory landscape analysis (multi-jurisdiction or evolving regimes)
• State-of-the-art reviews
• Pre-deal due diligence on a protocol or counterparty
• Literature reviews
• Multi-perspective investigation where conclusions matter to a downstream decision

Do NOT use:

• Simple lookups → a single-pass web-search skill or WebSearch
• Debugging → standard tools
• 1-2 search answers → a single-pass web-search skill
• Quick time-sensitive queries → a single-pass web-search skill
• Code review → a code-review skill
• Skill audit → skill-audit
• Pre-existing knowledge in your graph → recall first

Hard rules

1. Brain-first: ALWAYS run recall("<topic>") and check memory/Context/, knowledge/, your graph export before external retrieval. If we already know the answer or have prior research, surface that, don't regenerate.
2. Cite or omit: every load-bearing claim has an inline source citation. If you can't cite it, don't write it.
3. No fabricated citations: instant credibility loss + risk of poisoning the graph. Verify every URL resolves before locking.
4. Prose-first: bullet-soup is a tell of shallow research. Aim ≥80% prose.
5. Date everything: undated sources flagged explicitly, never silently treated as current.
6. Cluster independence: 3 sources tracing to the same primary = 1 source. Track upstream.
7. Surface high-materiality assumptions: in the Introduction, not buried.
8. CONFIRM-gate on knowledge-graph writes: per CLAUDE.md, knowledge-graph writes require the operator's confirmation. After Phase 7 (or 5 in std mode), present the structured output for approval BEFORE writing to the graph. Show: main report path + N finding pages + N source pages + N new entity stubs.
9. Heavy retrieval flags cost: ultradeep mode can burn $0.50+ in web-search auto-queries + browser-automation steps. Surface the estimated cost before running ultradeep.
10. Honest negatives: if research can't reach a confident answer, the report says so. Do NOT pad to look complete.

Out of scope

• Skill audit (use skill-audit)
• App-level OWASP / STRIDE (use a CSO skill)
• Single-fact lookups (use a single-pass web-search skill or WebSearch)
• Pre-existing knowledge retrieval (use recall)
• Code-context research (use a docs MCP such as mcp__plugin_context7_context7__query-docs)

Optional: a citation verifier

Consider adding a citation-verifier script (one you provide) that programmatically resolves every cited URL and flags dead links. Until you wire one in, do the URL-resolve check manually as part of the quality gates.

Source archive

The original SKILL.md is archived under sandbox/archive/skill-files/. Verdict during the audit: PARK + adapt. Adaptations made for this stack:

1. Output path: an Obsidian vault path → a knowledge-graph pages/Research___...md path
2. Frontmatter format: YAML → the graph's key:: value property block
3. Wiki-link namespacing: generic [[Entities/Name]] → namespaced [[Entity/Name]] matching the graph's namespace structure (Source/, Finding/, Entity/, Research/)
4. Install pattern: removed git clone main + cp -r (a Chain C violation). Methodology inlined here, scripts deferred.
5. Retrieval order: added web-search-primary + browser-automation-secondary + a brain-first preamble per the MCP stack
6. CONFIRM-gate: added hard rule 8. knowledge-graph writes pause for confirmation per CLAUDE.md
7. Trigger scoping: dropped bare "research" (collides with a single-pass skill); kept "deep research" / "literature review" / "comprehensive research" / etc.
8. Cost flagging: added hard rule 9 (estimate cost before ultradeep)
9. Dropped the Obsidian-Dataview-compatibility clause (irrelevant to the target graph)