Research

Research — L1 Base Deep Research

Standard research tier. Adds Planner (query decomposition) and per-source summarization on top of quick-research. This is the true foundation of the L1–L5 ladder — every higher level calls this one first, then layers on more.

Position in ladder: L1. Called by L2 (deep-research), L3, L4, L5 as the first stage.

When to Use

• "расскажи про X", "что такое Y", "как работает Z" — non-trivial but not strategic
• Technical overviews of a framework, library, concept
• "что нового в X в 2026"
• Default when user asks for research without specifying depth

Escalate to higher levels if:

• User wants comparison → /deep-research (L2) or /expert-research (L3)
• User wants strategic decision → /expert-research (L3)
• User wants academic rigor → /academic-research (L4)
• User wants a knowledge base → /ultra-research (L5)

Budget

• Time: ~5 min
• Tavily credits: ~20
• Sub-questions: 3
• Sources scraped: 10–15
• Output: structured report ~1000 words

Pipeline

1. CLARIFY     — skip if query is clear
2. PLAN        — decompose into 3 sub-questions
3. SEARCH      — parallel search for each sub-question
4. READ        — scrape 10-15 most relevant sources
5. SUMMARIZE   — per-source summary files
6. SYNTHESIZE  — structured report from summaries
7. SOURCES     — annotated bibliography

Artifacts Directory

Create a single session folder:

.firecrawl/research/<slug>/L1/
├── plan.md              # planner output (sub-questions)
├── sources/
│   ├── 01-<slug>.md     # full scrape
│   ├── 01-<slug>.sum.md # 300-500 word summary
│   ├── 02-...
│   └── ...
├── report.md            # final synthesized report
└── bibliography.md      # annotated sources

Use a URL-safe slug from the query (e.g., nextjs-edge-runtime).

Step 1: CLARIFY

If query is clear, skip. Otherwise ONE question max, then proceed.

Clear enough to proceed:

• "расскажи про Drizzle ORM"
• "как работает Tavily Research API"
• "что нового в Bun 2.0"

Needs clarification:

• "расскажи про auth" → ask: какая именно auth, для какого стека?
• "исследуй маркетплейсы" → ask: b2c/b2b, регион, вертикаль?

Step 2: PLAN (Planner decomposition)

Decompose the query into exactly 3 sub-questions that together cover the topic. Save to plan.md.

Decomposition heuristics:

• "Что такое X" → What, How, Why/When
  • What is X? (definition, components)
  • How does X work? (mechanism, architecture)
  • When/why to use X? (use cases, trade-offs, alternatives)
• "Как работает X" → Overview, Internals, Gotchas
• "Что нового в X" → Changes, Breaking, Migration
• "X vs Y" — use /deep-research instead, L1 is not for comparisons

Write plan.md:

# Research plan: <query>

**Session:** <slug>
**Level:** L1
**Created:** <date>

## Sub-questions
1. <subq1>
2. <subq2>
3. <subq3>

## Search strategy
- subq1 → [english query, russian query if applicable]
- subq2 → [...]
- subq3 → [...]

Step 3: SEARCH (parallel per sub-question)

Run ALL searches in parallel in ONE message (Bash calls + Tavily MCP calls together):

For each sub-question:

firecrawl search "<subq1 refined>" --limit 5 --json -o .firecrawl/research/<slug>/L1/search-1.json
firecrawl search "<subq2 refined>" --limit 5 --json -o .firecrawl/research/<slug>/L1/search-2.json
firecrawl search "<subq3 refined>" --limit 5 --json -o .firecrawl/research/<slug>/L1/search-3.json

Plus Tavily for each:

mcp__tavily__tavily_search with:
  query: <subq>
  max_results: 5
  search_depth: "advanced"

v0.2.2: persist Tavily results. Tavily goes through MCP — its output lives in conversation context by default, not on disk. After each mcp__tavily__tavily_search call, use the Write tool to save the returned JSON (or the full result object) to .firecrawl/research/$SLUG/L1/tavily-N.json where N matches the sub-question index. Without persistence, Tavily findings disappear after context compaction and can't be audited.

Search query refinement rules:

• Add year ("2026") for freshness-sensitive topics
• For Russian-language topics: one query in Russian + one in English
• For technical libraries: include library name + specific aspect
• Avoid generic queries — be specific

Step 4: READ (scrape 10–15 sources)

Combine results from all searches, deduplicate by URL, rank by:

1. Official docs (highest priority)
2. Authoritative sources (known tech blogs, framework maintainers)
3. Recent (2025–2026)
4. Community with real experience (HN, Reddit, GitHub discussions)

Pick 10–15 sources total (roughly 3–5 per sub-question).

Scrape ALL in parallel in one message using background tasks:

firecrawl scrape "<url1>" --only-main-content -o .firecrawl/research/<slug>/L1/sources/01-<slug>.md &
firecrawl scrape "<url2>" --only-main-content -o .firecrawl/research/<slug>/L1/sources/02-<slug>.md &
# ... up to 15
wait

Skip these:

• Paywalled content (NYT, WSJ, Medium paywall)
• SEO spam / thin affiliate pages
• Very old (pre-2023) unless historical context needed
• Translated machine-gen content

Step 5: SUMMARIZE (per-source, MANDATORY)

This is the key difference from L0. Each source gets its own summary file BEFORE synthesis.

For each scraped file, write a companion .sum.md file with this exact structure:

# Source 01: <title>
**URL:** <url>
**Type:** [official-docs | tech-blog | community | news | academic]
**Date:** <publication date if available>
**Quality:** [A | B | C]  <!-- A=primary/authoritative, B=solid secondary, C=supplementary -->
**Relevant to:** [subq1 | subq2 | subq3]

## Key facts
- Fact 1 (specific, citable)
- Fact 2
- ...

## Key quotes
> "exact quote 1"
> "exact quote 2"

## Notes
- Contradicts source X on Y (if applicable)
- Confirms source X on Z
- Gap: doesn't address W

Target: 300–500 words per summary. Extract facts, numbers, specific claims. Avoid generic statements.

Why this matters: synthesis step will read summaries, not raw scrapes. Keeps context clean, prevents hallucination, enables downstream levels (L2–L5) to work with structured input.

Step 6: SYNTHESIZE (report from summaries)

Read ALL .sum.md files. Produce report.md:

# <Topic Title>

**Query:** <original query>
**Level:** L1
**Sources:** <count>
**Generated:** <date>

## TL;DR
[2-3 sentences answering the core question directly]

## <Sub-question 1 as section heading>
[Answer grounded in sources. Inline citations [1], [2].]

## <Sub-question 2 as section heading>
[...]

## <Sub-question 3 as section heading>
[...]

## Key takeaways
- Point 1 [1, 3]
- Point 2 [2, 5]
- Point 3 [4]

## Open questions / gaps
- What the sources did NOT answer (honest list)
- Where sources disagreed

## Recommendation
[Concrete, opinionated recommendation if the query implies a decision. Skip if pure information query.]

Rules for synthesis:

• Organize by theme / sub-question, NOT by source
• Every factual claim MUST have [N] citation
• If sources conflict, note both positions explicitly
• Use specific numbers and quotes, not vague generalities
• Target 800–1500 words for the main body

Step 7: SOURCES (bibliography)

Write bibliography.md:

# Bibliography

1. **[Title]**(URL) — Quality: A | Type: official-docs
   Contribution: primary source for X, confirmed Y
2. **[Title]**(URL) — Quality: B | Type: tech-blog
   Contribution: real-world benchmarks for Z
...

🛑 FINAL VERIFICATION (before delivering the report)

Run this Bash block. If verification fails, fix the underlying problem — do NOT ship a broken L1 report (downstream L2/L3/L4/L5 depend on it being correct).

v0.4.0: verification logic now lives in the shared library at ~/.claude/scripts/lib/verify-research.sh (installed by scripts/install.sh). Skills call it instead of inlining 30+ lines of Bash. Single source of truth for citation regex, size thresholds, and source-pair counting. Fixing a check in one place updates every tier.

SLUG="<slug>"                  # replace with actual slug

# Locate the shared verification library.
# Prefer the installed copy; fall back to repo-local if running without install.sh.
VERIFY_LIB="$HOME/.claude/scripts/lib/verify-research.sh"
[ -f "$VERIFY_LIB" ] || VERIFY_LIB="scripts/lib/verify-research.sh"
[ -f "$VERIFY_LIB" ] || { echo "❌ verify-research.sh not found — run scripts/install.sh first"; exit 1; }

source "$VERIFY_LIB"
verify_l1 "$SLUG" || exit 1

Only deliver the report to the user after this prints ✅ L1 VERIFIED.

The verify_l1 function checks: plan exists; report ≥700 words; bibliography present; ≥10 scrape+summary pairs; each source ≥500 bytes; every [N] citation (including multi-cite [1, 3, 16]) maps to a bibliography entry.

Final output to user

1. Display report.md content in chat
2. Mention artifact location: Файлы в .firecrawl/research/<slug>/L1/
3. Offer next steps:

   Нужно глубже? /deep-research (L2) добавит reflection и проверку противоречий. Нужна строгая проверка? /expert-research (L3) добавит critic-агент и fact-check.

Rules

• Always planner first — never skip Step 2. This is what makes L1 ≠ L0.
• Always per-source summaries — never skip Step 5. Downstream levels depend on it.
• Parallel everything — searches and scrapes run concurrently in one message
• Cite everything — no claims without [N] reference
• Flag conflicts — explicit "source A says X, source B says Y"
• Russian / English — match user's query language in output
• No hallucinations — if sources don't answer it, say "not found in sources"
• Leave artifacts — keep files, don't clean up automatically (unlike L0)

Called from higher levels

When L2/L3/L4/L5 call this skill, they:

1. Execute the full L1 pipeline
2. Read report.md and all .sum.md files as input for their additional layers
3. Do NOT duplicate L1 work

This means L1 must produce clean, structured, self-contained output that downstream levels can build on.