/tech-search

name: tech-search
description: |
  Self-contained deep tech research. WebSearch + WebFetch + Haiku workers.
  Pipeline: Query > Decompose > Parallel Search (Haiku) > Evaluate > Synthesize > Document.
  Zero external dependencies. MCPs optional.
  Salva em docs/research/{YYYY-MM-DD}-{slug}/.

Tech Search

Self-contained deep research pipeline. Zero external dependencies.

Quick Start

/tech-search "React Server Components vs Client Components"

Activation

1. Parse query from `$ARGUMENTS` (or ask if not provided) 2. Execute 6-phase workflow 3. Save to `docs/research/{YYYY-MM-DD}-{slug}/`

**CRITICAL:**

• NEVER implement code. Redirect to @pm or @dev.
• NEVER write files outside `docs/research/`.

---

SKILL DEFINITION

skill:
  name: Tech Search
  id: tech-search

veto_conditions:
  - id: VETO_NO_RESULTS
    trigger: "ALL search waves return 0 results"
    action: "STOP + Report: 'No results found. Reformulate query or check connectivity.'"

  - id: VETO_IMPLEMENTATION_REQUEST
    trigger: "User asks to implement, code, create agent/skill, or deploy"
    action: "REDIRECT: 'Implementation is not my scope. Use @pm for prioritization or @dev for execution.'"
    keywords:
      - "implementa"
      - "cria o agent"
      - "cria a skill"
      - "faz o codigo"
      - "escreve o codigo"
      - "desenvolve"
      - "deploy"
      - "implement"
      - "build this"
      - "code this"

  - id: VETO_FORBIDDEN_PATH
    trigger: "Attempt to write outside docs/research/"
    action: "BLOCK + Error: 'Writing outside docs/research/ is forbidden.'"

constraints:
  forbidden_actions:
    - NEVER implement code, agents, skills, or production artifacts
    - NEVER create files outside docs/research/
    - NEVER write to .claude/agents/, .claude/skills/, squads/, app/, lib/

tool_hierarchy:
  search:
    1_preferred: "Exa MCP (mcp__exa__web_search_exa) - if available"
    2_fallback: "WebSearch (always available)"
    detection: "Try Exa first. If 401/429/503, set exa_available=false, use WebSearch."

  docs:
    1_preferred: "Context7 MCP (mcp__context7__resolve-library-id + query-docs) - if available"
    2_fallback: "WebSearch with 'site:{library}.dev docs' or 'site:{library}.io docs'"
    detection: "Try Context7 first. If fails, set context7_available=false."

  deep_read:
    only: "WebFetch with prompts/page-extract.md prompt"
    note: "No ETL, no Bash, no external scripts. Pure WebFetch."

  workers:
    type: "general-purpose"
    model: "haiku"
    max_parallel: 5
    max_deep_reads_per_worker: 3

workflow:
  phases:

    # ──────────────────────────────────────────────
    # PHASE 1: AUTO-CLARIFY
    # ──────────────────────────────────────────────
    1_auto_clarify:
      name: "Auto-Clarification"
      model_tier: "MAIN MODEL (inline)"
      description: |
        Pattern matching + technology detection on the user query.
        Determines if clarification is needed or can be skipped.

      execution: |
        1. Read user query (original text, unmodified)

        2. PATTERN MATCHING (case-insensitive):
           - Technical keywords: "code", "implement", "how to", "api", "bug",
             "error", "debug", "library", "sdk", "tutorial", "example"
             → inferred_context.focus = "technical"
           - Comparison keywords: "compare", "vs", "versus", "difference",
             "better", "alternative", "tradeoff", "pros and cons"
             → inferred_context.focus = "comparison"
           - Recency keywords: "latest", "new", "2024", "2025", "2026",
             "recent", "state of the art", "trending"
             → inferred_context.temporal = "recent"
             → Append current year to search queries

        3. TECHNOLOGY DETECTION (case-insensitive):
           Scan for known technologies:
           - Languages: JavaScript/JS, TypeScript/TS, Python, Java, Go, Rust, C#, Ruby, PHP
           - Frameworks: React, Next.js, Vue, Angular, Svelte, Express, FastAPI, Django, Flask
           - Databases: PostgreSQL, MySQL, MongoDB, Redis, Supabase, Firebase, Elasticsearch
           - AI/ML: LLM, RAG, LangChain, OpenAI, Claude, Anthropic, TensorFlow, PyTorch
           - Infra: Docker, Kubernetes, AWS, Vercel, GraphQL, REST, WebSocket
           → Collect into inferred_context.domain = [list]

        4. DECISION:
           - IF any pattern OR technology detected → skip clarification
           - IF nothing detected → ask ONE question:
             "Your query seems broad. What is the focus and technical context?"

      output: "inferred_context object {focus, temporal, domain, skip_clarification}"

    # ──────────────────────────────────────────────
    # PHASE 2: DECOMPOSE
    # ──────────────────────────────────────────────
    2_decompose:
      name: "Query Decomposition"
      model_tier: "MAIN MODEL"
      description: |
        Decomposes user query into 5-7 atomic, directly searchable sub-queries.
        Uses extended thinking for deeper analysis.

      execution: |
        ultrathink

        1. DEEP ANALYSIS (use extended thinking):
           - What are the REAL questions behind this query?
           - What would a domain expert want to know?
           - What gaps might standard searches miss?
           - What assumptions should be tested?

        2. GENERATE 5-7 sub-queries that:
           - Cover ORTHOGONAL angles (not overlapping)
           - Include at least one "devil's advocate" query
           - Include at least one "expert-level" query
           - Are directly searchable (not abstract)

        3. INCORPORATE inferred_context:
           - If focus=comparison → ensure queries cover both/all sides
           - If temporal=recent → add year constraints
           - If domain detected → scope queries to those technologies

        4. OUTPUT format:
           {
             "main_topic": "string",
             "sub_queries": ["query1", "query2", ...],
             "search_strategy": "parallel"
           }