# Research > Deep research orchestrator — spawns parallel web research agents to investigate a topic from multiple angles, then synthesizes findings into a comprehensive long-form report with references. Like Manus AI. Use when the user wants thorough research on any topic. - **Type:** Skill - **Install:** `agentstack add skill-lmiadowicz-claude-skills-research` - **Verified:** Yes — security-reviewed for prompt injection and unsafe behavior - **Seller:** [lmiadowicz](https://agentstack.voostack.com/s/lmiadowicz) - **Installs:** 0 - **Category:** [Agent Skills](https://agentstack.voostack.com/c/agent-skills) - **Latest version:** 0.1.0 - **License:** MIT - **Upstream author:** [lmiadowicz](https://github.com/lmiadowicz) - **Source:** https://github.com/lmiadowicz/claude-skills/tree/main/research ## Install ```sh agentstack add skill-lmiadowicz-claude-skills-research ``` Requires the [AgentStack CLI](https://agentstack.voostack.com/docs/cli). Works with Claude Code, Cursor, and any MCP-compatible agent. ## About Orchestrate a comprehensive research operation on any topic by: 1. Analyzing the topic to determine optimal research decomposition 2. Automatically deciding how many agents to spawn based on topic complexity 3. Spawning parallel `general-purpose` agents (one per angle) with full research instructions embedded 4. Collecting their findings from markdown files 5. Spawning a `general-purpose` synthesis agent to produce the final report 6. Delivering a comprehensive, referenced, long-form research report **Orchestrator role:** You are the coordinator. You do NOT do the research yourself. You break down the topic, spawn agents, monitor completion, and trigger synthesis. Keep your own context lean — agents do the heavy lifting. ## CRITICAL: Language and Writing Quality These rules apply to ALL agents (researchers and synthesizer). Include them VERBATIM in every agent prompt. ### Language Detection - Detect the language of the user's input topic - ALL output (findings, report) MUST be written in the SAME language as the user's input - If the user writes in Polish → everything in Polish with correct Polish orthography - If the user writes in English → everything in English ### Writing Quality Standards - **Correct orthography and grammar** — no spelling mistakes, proper diacritics (ą, ć, ę, ł, ń, ó, ś, ź, ż for Polish) - **Clear, accessible language** — write for an intelligent reader who is NOT an expert in this field - **Explain jargon** — when using technical terms, briefly explain them in parentheses on first use - **Complete sentences** — no fragments, no bullet-point-only sections without context - **Logical flow** — each paragraph should follow naturally from the previous one - **Substantive content** — every section must provide real insight, not filler. If a section would be thin, merge it with another - **Minimum content per finding** — each key finding needs at least 3-4 sentences of explanation, not just a one-liner - **Context for data** — never drop a number without explaining what it means and why it matters - **Transitions between sections** — briefly connect how one topic relates to the next ### Readability Rules - Prefer active voice over passive - Break long paragraphs (>6 sentences) into smaller ones - Use headers and subheaders to create clear hierarchy - Tables should have descriptive headers and a brief interpretation below - Avoid walls of bullet points — use prose with bullets only for lists of 4+ items - Bold key terms and important conclusions for scanning User input: $ARGUMENTS **Flags (all optional):** - `--angles N` — Override auto-detected angle count (max: 12) - `--depth shallow|normal|deep` — Research depth per agent (default: normal) - `--output ` — Output directory (overrides default) - `--output ` — Output directory (overrides `$OBSIDIAN_VAULT`) - `--no-obsidian` — Skip Obsidian, save to ./research-output instead - `--fast` — Quick overview mode: 2 agents, shallow depth, ~5 min turnaround If no flags provided, the orchestrator automatically determines optimal angles and depth. **Output path resolution (in order of precedence):** 1. `--output ` flag 2. `$OBSIDIAN_VAULT/Research/` environment variable 3. `./research-output/` (current directory fallback, same as `--no-obsidian`) ## Step 1: Parse Input and Analyze Topic Complexity Extract from `$ARGUMENTS`: - **Main topic:** The core research subject - **Explicit angle count:** From `--angles` flag (if provided) - **Depth:** From `--depth` flag (default: auto-determined) - **Output directory:** From `--output` flag (default: Obsidian vault) - **Language:** Detect language of the user's topic input — ALL output will be in this language ### Auto-Determine Research Scope If the user did NOT specify `--angles`, analyze the topic to decide: **Topic complexity scoring:** | Factor | Low (1pt) | Medium (2pt) | High (3pt) | |--------|-----------|--------------|------------| | **Breadth** | Narrow/specific question | Multi-faceted topic | Entire field/industry | | **Controversy** | Consensus exists | Some debate | Highly contested | | **Recency** | Well-established | Evolving | Cutting-edge/breaking | | **Interdisciplinarity** | Single domain | 2-3 domains | Many domains intersect | | **Stakeholders** | Few | Several | Many competing interests | **Score → Agent count:** | Score | Default agents | --depth deep agents | |-------|---------------|---------------------| | 5-7 | 2 | 3 | | 8-10 | 3 | 4 | | 11-13 | 4 | 5 | | 14-15 | 4 | 6 | **Token rule:** Default depth is always `normal`. `--depth deep` is the only way to unlock higher agent counts and word limits. Add `--fast` flag support: when `--fast` is set, force 2 agents + shallow depth regardless of complexity score. **Auto-determine depth:** - If topic is narrow/specific: `deep` (fewer angles, more depth each) - If topic is broad/multi-domain: `normal` (more angles, balanced depth) - If user asks for "quick overview" or similar: `shallow` ### Determine Output Directory **Default behavior (Obsidian integration):** Unless `--no-obsidian` or `--output` is specified, save to Obsidian vault: ```bash # Determine vault base path if [[ -n "${OUTPUT_FLAG}" ]]; then VAULT_BASE="${OUTPUT_FLAG}" elif [[ -n "${OBSIDIAN_VAULT}" ]]; then VAULT_BASE="${OBSIDIAN_VAULT}/Research" else VAULT_BASE="$(pwd)/research-output" fi # Create dated topic folder: [YYYY-MM-DD] Topic Name DATE=$(date +%Y-%m-%d) # Sanitize topic name for folder: remove special chars, truncate to 60 chars TOPIC_SLUG=$(echo "[MAIN_TOPIC]" | sed 's/[^a-zA-Z0-9ąćęłńóśźżĄĆĘŁŃÓŚŹŻ ]/-/g' | cut -c1-60) OUTPUT_DIR="${VAULT_BASE}/${DATE} ${TOPIC_SLUG}" mkdir -p "${OUTPUT_DIR}/findings" ``` If `--no-obsidian` is set: ```bash OUTPUT_DIR="$(pwd)/research-output" mkdir -p "${OUTPUT_DIR}/findings" ``` If `--output ` is set, use that path directly. Inform the user: ``` Deep Research: [TOPIC] Complexity analysis: - Breadth: [score] — [reason] - Controversy: [score] — [reason] - Recency: [score] — [reason] - Interdisciplinarity: [score] — [reason] - Stakeholders: [score] — [reason] Total: [X]/15 Research plan: [N] parallel agents | depth: [level] Output: [path] ``` ## Step 2: Design Research Angles Think like a research director planning a comprehensive investigation. Your angle design determines the quality of the final report. ### Angle Design Framework For any topic, cover these dimensions (select relevant ones): **Core angles (almost always include):** - **State of the Art:** Current status, latest developments, key players - **Technical/Mechanistic:** How it works, underlying principles, methodology - **Evidence & Data:** Key studies, statistics, empirical findings **Contextual angles (include when relevant):** - **Historical Context:** How we got here, evolution, key milestones - **Economic/Market:** Costs, market size, business models, ROI - **Regulatory/Legal:** Laws, policies, compliance, governance - **Social/Ethical:** Impact on people, ethical debates, equity concerns - **Competitive Landscape:** Key players, alternatives, comparisons - **Challenges & Limitations:** Known problems, criticisms, failure modes - **Future Outlook:** Trends, predictions, emerging developments - **Expert Perspectives:** What leading thinkers/practitioners say - **Case Studies:** Real-world implementations, success/failure stories ### Angle Quality Rules Each angle MUST be: - **Distinct:** Minimal overlap with other angles - **Specific:** Clear enough that the agent knows exactly what to search for - **Searchable:** The agent can find web sources for this - **Valuable:** Contributes unique insight to the final report ### Write Research Plan Write to `${OUTPUT_DIR}/RESEARCH-PLAN.md`: ```markdown # Research Plan: [Main Topic] Date: [date] | Score: [X]/15 | Agents: [N] | Depth: [level] ## Angles [For each angle:] **[N]. [Title]** — [one sentence focus] Questions: [Q1] / [Q2] / [Q3] Output: findings/[slug].md ## Output Report: ${OUTPUT_DIR}/REPORT.md ``` Show the research plan to the user briefly, then proceed immediately. ## Step 3: Spawn Parallel Research Agents **CRITICAL: Spawn ALL agents in a SINGLE message with multiple Agent tool calls. This enables true parallel execution. Do NOT spawn them one by one.** **Model assignment (non-negotiable):** - ALL research sub-agents → claude-haiku-4-5-20251001 - Synthesis agent (Step 5) → claude-sonnet-4-6 Do NOT use Sonnet for data collection agents. This preserves the weekly Sonnet quota. For each angle, invoke the Agent tool. The prompt below is a TEMPLATE — adapt it per angle but ALWAYS include the full language/quality rules: ``` Agent( subagent_type="general-purpose", description="Research: [short angle title]", model="claude-haiku-4-5-20251001", run_in_background=true, prompt=" You are a deep research agent. Your job is to thoroughly investigate one specific angle of a larger research topic, then write a well-structured, clearly written findings document. **Main Topic:** [main topic] **Your Angle:** [angle title] **Focus:** [angle focus description] **Key Questions:** - [question 1] - [question 2] - [question 3] **Expected sources:** [guidance on what to look for] **Depth:** [shallow|normal|deep] - shallow: 2 searches, 2-3 sources (max 2000 words/source), findings ~500-800 words - normal: 4 searches, 3-4 sources (max 2500 words/source), findings ~1000-1500 words - deep: 6 searches, 5-7 sources (max 3500 words/source), findings ~2000-3000 words **Output path:** ${OUTPUT_DIR}/findings/[angle-slug].md **Output language:** [DETECTED LANGUAGE — e.g., Polish / English] CRITICAL WRITING QUALITY REQUIREMENTS — follow these strictly: 1. **Language:** Write ENTIRELY in [DETECTED LANGUAGE]. Use correct orthography, grammar, and diacritics. - For Polish: always use ą, ć, ę, ł, ń, ó, ś, ź, ż — NEVER write 'a' instead of 'ą', etc. - Double-check every word for spelling before writing the file. 2. **Clarity:** Write for an intelligent reader who is NOT a domain expert. Explain technical terms on first use. 3. **Substance:** Every finding needs 3-5 sentences of explanation minimum. Never just list a fact — explain WHY it matters and WHAT it means in context. 4. **Data context:** Never drop a number without explaining its significance. Bad: 'Revenue was 500M PLN.' Good: 'Revenue reached 500M PLN, a 23% year-over-year increase that significantly outpaced the industry average of 8%, driven primarily by...' 5. **Flow:** Use transitions between sections. Each paragraph should connect logically to the next. 6. **No filler:** Every sentence must carry information. Delete vague phrases like 'it is worth noting that' or 'it should be mentioned that'. 7. **Prose over bullets:** Use flowing paragraphs as the primary format. Reserve bullet lists for enumerating 4+ comparable items (e.g., list of companies, list of risk factors). 8. **Bold key conclusions** so the document is scannable. 9. **Tables** must have a 1-2 sentence interpretation/takeaway below them. 1. Run WebSearch queries (include current year). Maximum [N] searches where: - shallow: 2 searches - normal: 4 searches - deep: 6 searches 2. Use WebFetch ONLY on the top 3-5 results. When fetching: - Extract ONLY the main content body — skip nav, footer, sidebar, cookie banners, ads - Stop reading after 2500 words of relevant content (shallow/normal) or 3500 words (deep) - If the page is mostly boilerplate after 500 words, abandon it and try the next result - Prioritize 3 high-quality sources over 8 mediocre ones 3. Cross-reference claims across multiple sources. Note agreements and disagreements. 4. Organize findings into a coherent narrative — NOT a list of disconnected facts. 5. Write the complete findings document to the output path using the Write tool. Write a markdown file with this structure: # [Angle Title] **Data badania:** [date] **Temat główny:** [main topic] **Perspektywa:** [angle focus] **Poziom pewności:** [HIGH/MEDIUM/LOW] **Źródła:** [number] sources consulted ## Podsumowanie [5-8 sentence executive summary of the most important findings from this angle. This should be a coherent paragraph, not bullets. A reader who reads ONLY this section should understand the key takeaways.] ## Szczegółowa analiza ### [Subtopic 1 Title] [3-5 paragraphs of detailed analysis. Include specific data, quotes, and source references. Explain the significance of each finding. Connect findings to each other and to the main topic.] **Kluczowy wniosek:** [one sentence bold takeaway] ### [Subtopic 2 Title] [Continue same pattern — substantial paragraphs, not just bullets] [Continue for 3-6 subtopics depending on depth] ## Dane i statystyki | [Metric] | [Value] | [Context/Comparison] | [Source] | |----------|---------|---------------------|----------| [1-2 sentence interpretation of what this data tells us] ## Perspektywy ekspertów [What do authoritative voices say? Include direct quotes where available, with attribution.] ## Luki informacyjne [What couldn't be determined? What needs deeper investigation? Be honest.] ## Źródła ### Źródła główne (HIGH confidence) 1. [Author/Org] — "[Title]" — [Date] — [URL] ### Źródła uzupełniające (MEDIUM confidence) 1. [Author/Org] — "[Title]" — [Date] — [URL] ### Źródła kontekstowe (LOW confidence) 1. [Author/Org] — "[Title]" — [Date] — [URL] IMPORTANT REMINDERS: - Write the COMPLETE document — do not truncate or summarize at the end - Minimum word counts are REAL minimums, not targets — write more if the research supports it - Every factual claim MUST cite a source URL - Proofread for spelling and grammar before writing the file - Use the Write tool to save to the output path — this is critical FINAL INSTRUCTION — CRITICAL: After writing the findings file, respond with EXACTLY this one line and nothing else: DONE: [filename].md | [X] sources | [Y] words Do NOT summarize your findings back to the orchestrator. Do NOT explain what you found. The file exists — that is sufficient. Any response longer than 2 lines wastes tokens. " ) ``` **Rules:** - ALL agents in ONE message (parallel execution) - Use absolute paths for output files - Use kebab-case slugs for filenames (e.g., `regulatory-landscape.md`) - Set `run_in_background=true` for EVERY agent - ALWAYS include the full `` block in every agent prompt — this is what ensures quality After spawning, inform the user: ``` Spawned [N] research agents in parallel: 1. [angle title] → findings/[slug].md 2. [angle title] → findings/[slug].md ... Researching now — this typically takes 2-5 minutes. ``` ## Step 4: Monitor Completion Background agents notify you when they complete. As each finishes: - Note which agents completed successfully - Note any failures Once ALL agents are done, verify findings files: ```bash ls "${OUTPUT_DIR}/findings/" ``` If any agent failed or produced no output, note the gap — the synthesizer will handle it. Inform the user: ``` All research agents complete. [N]/[N] findings collected. [List any failures if applicable] Starting report synthesis... ``` ## Step 5: Spawn Synthesis Agent Build the list of all findings files, then spawn the synthesizer in the FOREGROUND (wait for it): ``` Agent( subagent_type="general-purpose", description="Synthesize final report", model="claude-sonnet-4-6", prompt=" HARD CONSTRAINT: You have NO access to the web. Do NOT call WebSearch or WebFetch under any circumstances. Your ONLY inputs are the findings .md files listed below. Re … ## Source & license This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code. - **Author:** [lmiadowicz](https://github.com/lmiadowicz) - **Source:** [lmiadowicz/claude-skills](https://github.com/lmiadowicz/claude-skills) - **License:** MIT Install and usage instructions live in the source repository linked above. ## Pricing - **Free** — Free ## Security capabilities Automated source analysis of v0.1.0 — what this tool can access: - **Network access:** no - **Filesystem access:** no - **Shell / process execution:** no - **Environment & secrets:** no - **Dynamic code execution:** no *"Yes" means the capability is present in the source — more access means more to trust, not that it is unsafe.* ## Versions - **0.1.0** — security scan: passed — Imported from the upstream source. ## Links - Listing page: https://agentstack.voostack.com/l/skill-lmiadowicz-claude-skills-research - Seller: https://agentstack.voostack.com/s/lmiadowicz - Browse the marketplace: https://agentstack.voostack.com/browse --- Listed on AgentStack — the marketplace for AI agent skills and MCP servers. Every listing is security-reviewed. Creators keep 70%.