Academic Research

Academic Research — L4

Academic-grade research tier. Composes on top of L3 (/expert-research) by adding:

• Academic sources — arXiv, Google Scholar, SSRN, research.google, research papers
• Full multi-agent crew — Planner, Researcher × 2 (parallel), Critic, Editor
• Timeline / history analysis — how the topic evolved over time
• Methodology section — explicit description of research approach
• Cross-language search — EN + RU + domain-specific languages
• Quality-rated bibliography — A/B/C/D ratings per source
• Open questions section — what the literature hasn't answered

Position in ladder: L4. Calls L3 (which chains through L2 and L1). Called by L5.

When to Use

• Scientific/technical deep-dives: "как работают LLM router-агенты"
• Research-grade investigations needing peer-reviewed backing
• Topics with substantial academic literature (ML/AI, distributed systems, algorithms)
• When user says "академический", "научный", "peer-reviewed", "с научной точки зрения"
• Complex topics where vendor-driven sources aren't enough

Escalate to:

• /ultra-research (L5) — when you want a full knowledge base + playbooks + memory sync

Budget

• Time: ~40 min total (L1: 5 + L2: 7 + L3: 8 + L4: 20)
• Tavily credits: ~400
• Sub-questions: 12
• Sources: 80–120 (including 15–25 academic)
• Output: 5000+ word academic report + annotated bibliography + timeline + methodology

Pipeline

STAGE 1: Execute /expert-research (L3)
    → L1/ + L2/ + L3/ artifacts

STAGE 2: L4 Academic Layer
  1. ACADEMIC PLANNING           — scope the academic angle
  2. MULTI-AGENT CREW SETUP      — spawn crew in parallel
     ├─ Planner agent            — generates 4 new academic sub-qs
     ├─ Researcher A (parallel)  — primary academic sources
     ├─ Researcher B (parallel)  — cross-domain / adjacent fields
     ├─ Critic (after R A+B)     — adversarial review of academic layer
     └─ Editor (last)            — final synthesis pass
  3. ACADEMIC SOURCE HARVESTING  — arXiv, Scholar, SSRN
  4. CROSS-LANGUAGE SEARCH       — EN + RU + other relevant
  5. TIMELINE CONSTRUCTION       — chronological development
  6. METHODOLOGY DOCUMENTATION   — how research was done
  7. QUALITY-RATED BIBLIOGRAPHY  — A/B/C/D per source
  8. ACADEMIC SYNTHESIS          — formal structure

Artifacts Directory

.firecrawl/research/<slug>/
├── L1/ L2/ L3/                  # from lower levels
└── L4/
    ├── academic-plan.md
    ├── crew/
    │   ├── planner-output.md
    │   ├── researcher-a-output.md
    │   ├── researcher-b-output.md
    │   ├── critic-output.md
    │   └── editor-output.md
    ├── sources/
    │   ├── academic/            # arXiv, Scholar papers
    │   │   ├── 50-<paper>.md
    │   │   └── 50-<paper>.sum.md
    │   └── secondary/
    │       ├── 70-...
    │       └── ...
    ├── timeline.md
    ├── methodology.md
    ├── open-questions.md
    ├── report.md                # 5000+ word academic report
    ├── executive-summary.md     # updated from L3
    ├── bibliography.md          # A/B/C/D rated
    └── report.pdf

Stage 1: Execute L3

Invoke expert-research skill. Use --go flag semantic (no plan approval prompt — we're already committed at L4):

Skill: expert-research
Args: <query> (note: running as L4 foundation, skip plan approval)

Wait for L3 completion. Verify L3 artifacts exist before proceeding.

Stage 1b: Exa as primary academic source (optional, v0.5.0+)

Biggest L4 upgrade in v0.5.0. Previously L4 used Tavily with "academic" keywords — hit-or-miss. Exa has a native research paper category indexing arXiv, paperswithcode, and other academic sources.

If Exa MCP is installed, use web_search_advanced_exa with category: "research paper" as the primary academic discovery channel. Tavily remains as complementary coverage (non-academic sources, recent industry reports).

# For each academic sub-question generated by the Planner crew member:
mcp__exa__web_search_advanced_exa with:
  query: <academic sub-question>
  category: "research paper"
  type: "auto"
  num_results: 15
  contents: { text: { max_characters: 20000 } }

Save each response to .firecrawl/research/$SLUG/L4/exa-academic-<subq-num>.json. Ranking, scoring, and full text are returned in one call — no separate scrape needed for most papers.

Quality grading: Exa results with category: "research paper" automatically qualify as Quality A (primary) in the bibliography — they're peer-reviewed or preprint-server-hosted. Web/news results stay B/C as before.

If Exa unavailable: skill falls back to pre-v0.5 behavior (Tavily + manual arXiv scraping via Firecrawl) — everything still works, just less academic coverage per time-budget.

Stage 1a: Codex as Researcher C (optional, added v0.2)

Fault-tolerant — if Codex is unavailable, the crew runs with 2 researchers instead of 3. Skill continues.

As part of the multi-agent crew (Step 2 below), Researcher C = Codex CLI (GPT-5.4 with live search). This gives the crew a genuinely different model with a different search index, complementing Researcher A (Claude general web) and Researcher B (Claude academic).

Launch Researcher C as soon as L3 finishes, in parallel with the crew setup:

CODEX_HELPER="$HOME/.claude/scripts/codex-research.sh"
[ -x "$CODEX_HELPER" ] || CODEX_HELPER="scripts/codex-research.sh"

if [ -x "$CODEX_HELPER" ]; then
    bash "$CODEX_HELPER" 360 \
        ".firecrawl/research/$SLUG/L4/crew/researcher-c-output.md" \
        "You are Researcher C in an academic-grade research crew. Topic: <ORIGINAL QUERY>. Researchers A and B are covering general web and academic databases. Your role: find ADJACENT-field sources and independent evaluations that the others might miss. Look at conference talks, industry reports, niche communities, non-English sources where relevant. Return 10-15 findings with URLs and brief context (≤1500 words). Include source type (industry-report | conference | community | trade-publication | other)." &
    CODEX_C_PID=$!
else
    CODEX_C_PID=""
fi

Wait for Researcher C after the crew's Critic step completes. If it produced output, incorporate its findings in Step 8 (Academic Synthesis) with (Researcher C / cross-model) citations. If not (status file explains why), the crew proceeds as 2-researcher.

Stage 2: L4 Academic Layer

Step 2.1: ACADEMIC PLANNING

Read L3/report.md. Identify which parts have thin academic backing. Generate academic research scope:

Write L4/academic-plan.md:

# Academic Research Plan

## Current academic coverage in L3
- <what academic sources L3 already used>
- <gaps in academic grounding>

## Academic sub-questions (4 new)
9. What does peer-reviewed literature say about X?
10. What's the academic history/evolution of X?
11. What are the theoretical foundations of X?
12. What open research problems exist in X?

## Academic sources to target
- arXiv: [specific categories e.g. cs.LG, cs.CL]
- Google Scholar: [key author names, seminal papers]
- SSRN: [if applicable]
- Research.google: [if applicable]
- University CS departments: [Stanford, MIT, CMU relevant]

Step 2.2: MULTI-AGENT CREW SETUP

Spawn a crew of sub-agents. Use Agent tool with parallel dispatch for Researcher A and Researcher B.

Important: use run_in_background: true for parallel researchers so they work simultaneously.

Planner Agent (synchronous, fast)

Agent(subagent_type="general-purpose",
      description="Academic planner for L4",
      prompt="""
You are an academic research planner.

Read: .firecrawl/research/<slug>/L3/report.md
And: .firecrawl/research/<slug>/L4/academic-plan.md

Your job: refine the 4 academic sub-questions into specific, searchable queries.

For each sub-question produce:
- 3 specific search queries (mix of simple and advanced)
- Target source types (arXiv category, Scholar keywords)
- Expected difficulty (easy / medium / hard to find)
- Why this sub-question matters for the main topic

Write output to: .firecrawl/research/<slug>/L4/crew/planner-output.md
""")

Researcher A (parallel, background)

Agent(subagent_type="general-purpose",
      description="Academic researcher A — primary sources",
      run_in_background=true,
      prompt="""
You are academic researcher A, focused on PRIMARY academic sources.

Read: .firecrawl/research/<slug>/L4/crew/planner-output.md
Your targets: arXiv papers, peer-reviewed journals, conference proceedings.

For each of the 4 sub-questions, do:
1. Search arXiv via Tavily or Firecrawl with "site:arxiv.org <query>"
2. Search Google Scholar via Firecrawl with "site:scholar.google.com <query>"
3. Scrape top 3-5 papers per sub-question
4. Write summary .sum.md files to .firecrawl/research/<slug>/L4/sources/academic/
5. Use numbering starting from 50

For each paper summary include:
- Title, authors, year, venue
- Abstract (from paper)
- Key findings
- Methodology used
- Limitations stated by authors
- Relevance to our sub-question
- Quality grade: A (top venue, cited) | B (decent venue) | C (preprint, limited citations)

Finally write: .firecrawl/research/<slug>/L4/crew/researcher-a-output.md
with:
- List of papers found (with quality grades)
- Key findings across papers
- Consensus / conflicts in the literature
- Gaps identified
""")

Researcher B (parallel, background)

Agent(subagent_type="general-purpose",
      description="Academic researcher B — cross-domain",
      run_in_background=true,
      prompt="""
You are academic researcher B, focused on CROSS-DOMAIN and ADJACENT-FIELD sources.

Read: .firecrawl/research/<slug>/L4/crew/planner-output.md

Your targets:
- Adjacent fields: if main topic is ML, look at stats/optimization/systems angles
- Cross-language sources: Russian, Chinese, French academic work if applicable
- Historical/foundational papers: seminal works that shaped the field
- Cross-disciplinary applications: how the topic is used in other domains

Search targets:
- Russian academic: cyberleninka.ru, elibrary.ru, habr.com research posts
- Classic papers: Google Scholar citation counts, ACM Digital Library
- Adjacent fields: different arXiv categories than Researcher A

For each source write .sum.md to .firecrawl/research/<slug>/L4/sources/secondary/
Number starting from 70.

Finally write: .firecrawl/research/<slug>/L4/crew/researcher-b-output.md
with:
- Cross-domain findings
- Cross-language findings (translated summaries if non-English)
- Historical context (which papers started the field, key evolutions)
- Adjacent-field insights we might have missed
""")

Wait for BOTH Researcher A and Researcher B to complete before proceeding.

Critic Agent (synchronous, after researchers)

Agent(subagent_type="general-purpose",
      description="Academic critic — L4 layer review",
      prompt="""
You are an academic critic. Review the L4 academic layer.

Read:
- .firecrawl/research/<slug>/L3/report.md
- .firecrawl/research/<slug>/L4/crew/researcher-a-output.md
- .firecrawl/research/<slug>/L4/crew/researcher-b-output.md
- Sample sources from L4/sources/academic/*.sum.md

Review questions:
1. Are the academic sources actually high-quality or just preprints?
2. Do Researcher A's findings contradict the L3 report's conclusions? If so, which is right?
3. Are there missing seminal papers or key authors?
4. Is the academic framing honest, or is it cherry-picking to support L3's conclusions?
5. What's the academic consensus vs minority views? Are minority views represented?
6. Are there methodological concerns with cited papers that should be flagged?

Write findings to: .firecrawl/research/<slug>/L4/crew/critic-output.md

Be harsh. Academic research needs academic rigor — not marketing.
""")

Editor Agent (last, after critic)

Agent(subagent_type="general-purpose",
      description="Academic editor — final synthesis",
      prompt="""
You are the editor. Your job: synthesize everything into a formal academic report.

Read ALL L1-L4 artifacts:
- L1/report.md, L2/report.md, L3/report.md
- L4/crew/researcher-a-output.md, researcher-b-output.md, critic-output.md
- L4/sources/academic/*.sum.md, L4/sources/secondary/*.sum.md
- L3/fact-check.md, L2/contradictions.md

Produce .firecrawl/research/<slug>/L4/crew/editor-output.md with:

## Consolidated findings across all 4 layers
## Where academic sources override popular-media sources
## Where minority academic views deserve mention
## Proposed structure for the final L4 report
## Key claims to emphasize with academic backing
## Claims to downgrade because academic literature contradicts them
""")

Step 2.3: TIMELINE CONSTRUCTION

Read all summaries from L1–L4. Build a chronological timeline of the topic.

Write L4/timeline.md:

# Timeline: <topic>

## Pre-<year> — Origins
- <year>: <event> — <citation>
- ...

## <year>-<year> — Early development
- ...

## <year>-<year> — Maturation
- ...

## <year>-2026 — Current state
- ...

## What's next (speculation based on recent papers)
- ...

Step 2.4: METHODOLOGY DOCUMENTATION

Write L4/methodology.md — transparency about how this research was conducted:

# Research Methodology

## Levels executed
- L1 (base): planner + 10-15 sources + summaries
- L2 (deep): reflection + contradiction detection + 10-15 additional sources
- L3 (expert): 3-angle search + fact-check + critic + 10-15 more sources
- L4 (academic): multi-agent crew + 15-25 academic sources

## Agents used (L4)
- Planner, Researcher A, Researcher B, Critic, Editor

## Source types and counts
| Type | Count | Quality A | Quality B | Quality C | Quality D |
|---|---|---|---|---|---|
| Official docs | N | ... |
| Academic papers | N | ... |
| Tech blogs | N | ... |
| Community | N | ... |

## Search engines used
- Firecrawl, Tavily
- arXiv.org (via site-specific search)
- Google Scholar (via site-specific search)

## Languages
- English: primary
- Russian: <count> sources
- Other: <count>

## Biases acknowledged
- <known limitations of this approach>
- <sources we couldn't access>
- <topics we couldn't verify independently>

## Reproducibility
- All source URLs listed in bibliography.md
- All search queries logged in */plan.md files
- Timestamp: <date>

Step 2.5: QUALITY-RATED BIBLIOGRAPHY

Re-grade ALL sources (L1 + L2 + L3 + L4) into A/B/C/D:

• A — Primary source (official docs, peer-reviewed paper in top venue, original research)
• B — Authoritative secondary (established tech blog, conference paper, industry report)
• C — Supplementary (community discussion, preprint, developer blog)
• D — Unreliable (SEO spam, low-quality aggregator, unsourced claims) — should rarely be cited

Write L4/bibliography.md:

# Annotated Bibliography (L4)

## Quality A sources (<count>)
1. **[Title]**(URL) — Authors, Year, Venue
   Type: peer-reviewed paper | official docs | ...
   Contribution: [what it adds to this research]
   Quality: A — [reason: top venue, X citations, authoritative]

## Quality B sources (<count>)
...

## Quality C sources (<count>)
...

## Quality D sources (<count>, flagged)
... [if any, explain why retained]

Step 2.6: OPEN QUESTIONS

Write L4/open-questions.md:

# Open Research Questions

Based on the academic literature surveyed, these are questions without a clear answer:

## Unresolved in the literature
- <question 1>: <why it's open, which papers touch it>
- ...

## Contested
- <question 2>: <camps, their positions>

## Implied by L4 findings but not directly studied
- ...

## What would resolve them
- [what kind of research / data / experiment would answer]

Step 2.7: ACADEMIC SYNTHESIS

Produce L4/report.md. This supersedes L3/report.md.

Academic-style structure:

# <Topic Title>: An Academic Review

**Query:** <original>
**Level:** L4 (Academic)
**Sources:** <total>, of which <academic count> peer-reviewed
**Generated:** <date>

## Abstract
[200-word academic abstract: background, method, findings, conclusion]

## 1. Introduction
### 1.1 Background
### 1.2 Scope of this review
### 1.3 Methodology (brief; see methodology.md for details)

## 2. Historical Development
[Based on timeline.md — chronological evolution with citations]

## 3. Theoretical Foundations
[If applicable — underlying theory, formalism, mathematical background]

## 4. Current State of the Art
### 4.1 [Sub-topic 1]
### 4.2 [Sub-topic 2]
### 4.3 [Sub-topic 3]
### 4.4 [Sub-topic 4]

## 5. Critical Analysis
### 5.1 Strengths
### 5.2 Limitations
### 5.3 Contested claims
### 5.4 Minority views worth considering

## 6. Cross-Domain Connections
[From Researcher B's findings]

## 7. Practical Implications
[How academic findings translate to practice — bridge to L3 practical content]

## 8. Open Questions
[Summary pointing to open-questions.md]

## 9. Conclusion

## References
[See bibliography.md for full annotated list]

## Appendices
- A. Search queries used (from all plan.md files)
- B. Timeline (see timeline.md)
- C. Methodology (see methodology.md)

Target length: 5000–7000 words.

Step 2.8: PDF Export

cd .firecrawl/research/<slug>/L4/
pandoc report.md -o report.pdf \
  --pdf-engine=xelatex \
  --toc \
  --number-sections \
  -V geometry:margin=1in \
  -V fontsize=11pt 2>/dev/null || \
pandoc report.md -o report.html --toc

Final Output

1. Show user the Abstract + Conclusion sections in chat
2. Stats:

   📊 L4 stats: sources ( academic, quality: A: B: C:) 🤖 Crew: 5 agents (Planner + 2×Researcher + Critic + Editor) 📄 Report: words, academic style

3. Artifacts:

   .firecrawl/research/<slug>/L4/ — report.md, report.pdf, timeline.md, methodology.md, open-questions.md, bibliography.md

4. Escalation:

   Нужна полная база знаний с playbook'ами и памятью? /ultra-research (L5) — финальный уровень.

Rules

• Always call L3 first — never inline L3 logic
• Crew runs in parallel — Researcher A and B use run_in_background: true
• Academic sources are non-negotiable — target 15+ peer-reviewed/preprint papers
• Quality grading is honest — most blog posts are B or C, not A
• Timeline is grounded in sources — every date needs citation
• Methodology is transparent — we explain what we did and didn't do
• Cross-language matters — at least attempt Russian sources when topic allows
• Preserve all lower-level artifacts

Called from L5

When L5 calls this skill:

• Run full L4 pipeline
• L5 reads L4/report.md, L4/timeline.md, L4/open-questions.md, L4/bibliography.md as foundation