--- name: research description: "Researches any topic from the last 30 days by dispatching 6-10 parallel sub-agents across community discussions and official sources. Use when user wants deep research, topic analysis, community sentiment, or asks 'what's new with X'." argument-hint: "[topic] for [tool]" or "[topic]" or "[topic] deep" context: fork agent: Explore --- # research: Deep Research Any Topic from the Last 30 Days Research ANY topic across Reddit, X, community forums, official docs, academic papers, and industry publications. Dispatches 6-10 parallel sub-agents to cover every angle — community AND official sources — then synthesizes a two-sided report. Use cases: - **Prompting**: "photorealistic people in Nano Banana Pro" — learn techniques from community tips AND official guides - **Recommendations**: "best Claude Code skills" — get specific names from community + official feature comparisons - **News**: "what's happening with OpenAI" — community reactions + official announcements - **General**: any topic — understand what the community says AND what officials report Input: `\/research best Claude Code skills` Output: Two-sided report with: - Community: "Most mentioned: \/commit (5x, r\/ClaudeAI + HN), remotion skill (4x, X + Dev.to)..." - Official: "Anthropic docs highlight \/pr, \/review as recommended workflows..." - Cross-reference: 3 aligned, 1 divergent, 1 gap - Stats: 12 Reddit threads (847 upvotes) + 8 X posts + 6 community + 4 official sources Then waits for user's vision to write a tailored prompt. <\/example> Edge case — Web-only mode (no API keys): SCRIPT_DATA returns "Mode: web-only". Sub-agents do ALL discovery work. Display omits Reddit\/X stats lines. Report still delivers full two-sided analysis. <\/example> Edge case — Agent failure: C4 returns empty or errors. Log "C4: dev communities — no results". Continue with C1-C3 data. Note gap in stats footer: "1 agent failed". <\/example> Edge case — Sources disagree: Community (C2): "Tool A recommended (12 upvotes on HN)" Official (O1): "Tool B is the documented approach" Cross-reference: Mark as "Divergent" with both views attributed. <\/example> ## Parse User Intent Extract four variables from user input before proceeding: | Variable | Extract | Example | |----------|---------|---------| | `TOPIC` | What they want to learn about | "web app mockups", "Claude Code skills" | | `TARGET_TOOL` | Where they'll use prompts (or "unknown") | "Nano Banana Pro", "Midjourney" | | `QUERY_TYPE` | PROMPTING \| RECOMMENDATIONS \| NEWS \| GENERAL | Auto-detect from phrasing | | `DEPTH` | `--quick` (6 agents) \| default (8) \| `--deep` (10) | Flag in user input | See `references\/intent_parsing.md` for query type definitions, detection patterns, and variable storage rules. **Do NOT ask about target tool before research.** If unspecified, ask AFTER showing results. State: TOPIC, TARGET_TOOL, QUERY_TYPE, DEPTH are now defined. Next: Determine API key availability to select script mode. <\/context> ## Setup Check The Python script works in three modes based on available API keys: 1. **Full Mode** (both keys): Reddit + X with real engagement metrics 2. **Partial Mode** (one key): Reddit-only or X-only 3. **Web-Only Mode** (no keys): Script provides no data, sub-agents do all the work API keys are OPTIONAL. The skill always dispatches sub-agents regardless. Determine mode quickly — don't block on this. ## MCP Tool Detection Use `ToolSearch` with query `"brave search"` to check for Brave MCP tools. Record as `AVAILABLE_MCP_TOOLS`. Construct the `MCP_TOOLS` instruction block embedded into every sub-agent prompt: - **Brave available**: Use `brave_web_search` and `brave_news_search` as PRIMARY tools; fall back to `WebSearch` only on errors - **No MCP tools**: Use `WebSearch` for all queries State: MCP_TOOLS instruction block is defined. Script runs synchronously. Output: SCRIPT_DATA (Reddit\/X results with engagement metrics, or empty if web-only mode). <\/context> ## Phase 1: Run Python Script (Reddit + X) Run the research script synchronously — it provides Reddit\/X data with real engagement metrics that sub-agents cannot replicate. ```bash python3 ~\/.claude\/skills\/research\/scripts\/last30days.py "$TOPIC" --emit=compact 2>&1 ``` The `$DEPTH` flag maps to: `--quick` -> pass `--quick`; default -> no flag; `--deep` -> pass `--deep`. Store the output as `SCRIPT_DATA`. Check mode from output: - **"Mode: both"** \/ **"Mode: reddit-only"** \/ **"Mode: x-only"**: Script found data - **"Mode: web-only"**: No API keys, sub-agents provide all data Do NOT stop or warn if web-only. Proceed to Phase 2. State: SCRIPT_DATA collected. All agents launched in ONE message for parallelism. Dependencies: MCP_TOOLS block embedded in every sub-agent prompt. <\/context> ## Phase 2: Sub-Agent Dispatch **All agents are launched in a SINGLE message with multiple Task tool calls for maximum parallelism.** Refer to `references\/subagent_prompts.md` for prompt templates and `references\/source_categories.md` for source taxonomy. ### Agent Allocation Dispatch community agents (C1-C5) and official agents (O1-O5) in parallel. Agent count scales with depth: - `--quick`: 6 (3C + 3O) | default: 8 (4C + 4O) | `--deep`: 10 (5C + 5O) Build each prompt from templates in `references\/subagent_prompts.md`, filling: `{TOPIC}`, `{QUERY_TYPE}`, `{FOCUS}`, `{QUERIES}`, `{DATE_FROM}`, `{MCP_TOOLS}`. Use `subagent_type: "general-purpose"`. See `references\/agent_allocation.md` for full agent roles, focus areas, and dispatch pattern. **DISPATCH ALL AGENTS IN A SINGLE MESSAGE.** Do not dispatch sequentially. ## Phase 3: Collect Results After dispatching, collect results from all agents: 1. Call `TaskOutput` for each dispatched agent 2. Organize into: `COMMUNITY_FINDINGS` (C1-C5) and `OFFICIAL_FINDINGS` (O1-O5) 3. **Graceful failure**: If an agent fails or returns empty, log which agent failed, continue with remaining results, note the gap in the final report. Do NOT retry. State: SCRIPT_DATA + COMMUNITY_FINDINGS + OFFICIAL_FINDINGS all collected. Mode: Synthesis — ground in actual research, not pre-existing knowledge. <\/context> ## Phase 4: Judge Synthesis Synthesize all findings (SCRIPT_DATA + COMMUNITY_FINDINGS + OFFICIAL_FINDINGS) into a coherent two-sided report. ### Weighting | Source | Weight | Why | |--------|--------|-----| | Reddit (from script) | HIGHEST | Real upvotes + comments = proven engagement | | X (from script) | HIGHEST | Real likes + reposts = proven engagement | | HN \/ Lobsters | HIGH | Voting system = community curation | | Official docs | HIGH | Authoritative, primary source | | Academic papers | HIGH | Peer-reviewed | | Industry publications | MEDIUM | Expert but potentially biased | | Expert blogs | MEDIUM-LOW | Individual perspective | | Misc community | MEDIUM-LOW | Volume varies | ### Cross-Reference Analysis Identify 3-5 topics where community and official sources can be compared: - **Aligned**: Both sides agree - **Divergent**: Community says one thing, officials say another - **Gap**: One side has information the other lacks ## Internalize the Research **Ground your synthesis in ACTUAL research content, not pre-existing knowledge.** Read all agent outputs carefully — pay attention to exact names, specific insights, and real engagement numbers. ### If QUERY_TYPE = RECOMMENDATIONS Extract SPECIFIC NAMES from ALL sources (script + community + official agents). Count mentions across sources, note which sources recommend each, list by popularity. BAD: "Skills are powerful. Keep them under 500 lines." GOOD: "Most mentioned: \/commit (5 mentions, r\/ClaudeAI + HN), remotion skill (4x, X + Dev.to), git-worktree (3x, GitHub Discussions). Official docs highlight \/pr as recommended." <\/example> ### For All Query Types From the ACTUAL RESEARCH OUTPUT, identify: - **PROMPT FORMAT** — Does research recommend JSON, structured params, natural language, keywords? - Top 3-5 patterns\/techniques that appeared across multiple sources - Specific keywords, structures, or approaches mentioned BY THE SOURCES - Common pitfalls mentioned BY THE SOURCES **If research says "use JSON prompts" or "structured prompts", you MUST deliver prompts in that format later.** **SELF-CHECK**: Re-read your synthesis before displaying. Does it match what the research ACTUALLY says? If you catch yourself projecting your own knowledge, rewrite it. ## Display Two-Sided Report Refer to `references\/output_format.md` for the full template. Do NOT output any "Sources:" lists. ### 1. What the Community Says Synthesize SCRIPT_DATA (Reddit\/X) + COMMUNITY_FINDINGS (C1-C5): If RECOMMENDATIONS: Most Mentioned: 1. [Specific name] - mentioned {n}x (r\/sub, HN, @handle, blog.com) 2. [Specific name] - mentioned {n}x (sources) 3. [Specific name] - mentioned {n}x (sources) Notable mentions: [other specific things with 1-2 mentions] <\/example> If PROMPTING\/NEWS\/GENERAL: What the community is saying: [2-4 sentences synthesizing key insights FROM THE ACTUAL RESEARCH OUTPUT.] Key patterns: 1. [Pattern from research] 2. [Pattern from research] 3. [Pattern from research] <\/example> ### 2. What the Official Sources Say Synthesize OFFICIAL_FINDINGS (O1-O5): - Key findings with authority attribution - Recent official changes with dates - Gaps in official coverage ### 3. Where They Agree and Disagree Cross-reference table (3-5 rows): | Topic | Community View | Official Position | Status | |-------|---------------|-------------------|--------| | [aspect] | [what community says] | [what officials say] | Aligned \/ Divergent \/ Gap | ### 4. Stats Footer Display real numbers from the research: ``` All agents reported back! |- Reddit: {n} threads | {upvotes} upvotes | {comments} comments |- X: {n} posts | {likes} likes | {reposts} reposts |- Community web: {n} sources (HN, forums, blogs) |- Official web: {n} sources (docs, papers, reports) |- Agents dispatched: {total} ({C}C + {O}O) |- Cross-reference: {agree} aligned, {disagree} divergent, {gaps} gaps ``` If web-only mode, omit Reddit\/X lines and add the API key hint from `references\/output_format.md`. ### 5. Invitation ``` Share your vision for what you want to create and I'll write a thoughtful prompt you can copy-paste directly into {TARGET_TOOL}. ``` **Use real numbers from the research output.** Patterns should be actual insights, not generic advice. **If TARGET_TOOL is still unknown after showing results**, ask now: ``` What tool will you use these prompts with? Options: 1. [Most relevant tool based on research] 2. Nano Banana Pro (image generation) 3. ChatGPT \/ Claude (text\/code) 4. Other (tell me) ``` After displaying the report and invitation, WAIT for the user to respond. ## Prompt Generation When the user shares their vision, write ONE tailored prompt using expertise from BOTH community and official sources. **Critical**: Match the FORMAT the research recommends (JSON, structured params, natural language, keywords). See `references\/prompt_generation.md` for the full prompt writing protocol, quality checklist, and output footer templates. ## Context Memory After research completes, retain topic expertise for follow-up questions. See `references\/context_memory.md` for full context retention protocol. <\/instructions>