team-shinchan-research

name	team-shinchan:research
description	Use when you need web research, documentation lookup, or knowledge gathering.
user-invocable	false

EXECUTE IMMEDIATELY

Step 1: Validate Input

If args is empty or only whitespace:
  Ask user: "What would you like to research?"
  STOP and wait for user response

If args length > 2000 characters:
  Truncate to 2000 characters
  Warn user: "Request was truncated to 2000 characters"

Step 1b: Detect Mode

Parse mode from args:

• If args starts with youtube or article or auto : extract mode as first word, remainder is the URL/query.
• If no mode keyword: mode = auto for URLs (args starts with http), else mode = search (standard web search).

Set target = the URL/query: the remainder after the mode keyword, or the whole args when there is no mode keyword. (target is referenced by the Step 2A prompt for every mode.)

Modes youtube, article, auto require a URL. If mode detected but no URL found, ask: "Please provide a URL for {mode} extraction."

Step 1c: Narrow vs Breadth

• Narrow (single source): mode youtube / article / auto (one URL), OR a search for a single specific fact/definition ("what is X", "X's default port", one API detail).
• Breadth (multi-source): a search that is comparative or survey-shaped — triggers include "compare", "vs", "survey", "which … should we", "options for", "best practices for", "investigate", "pros and cons", or anything that benefits from several independent sources.

Route: narrow → Step 2A (single Masumi). breadth → Step 2B (orchestrator-worker fan-out).

Step 2A: Narrow research (single Masumi)

Task(
  subagent_type="team-shinchan:masumi",
  model="sonnet",
  prompt=`/team-shinchan:research has been invoked.

## Research Request

Mode: ${mode}  <!-- youtube | article | auto | search -->
URL/Query: ${target}

Conduct thorough research and provide:

| Section | Content |
|---------|---------|
| Key Findings | Main discoveries with sources |
| Documentation | Relevant docs and reference links |
| Best Practices | Recommended approaches |
| Caveats | Potential concerns or limitations |

User request: ${args || '(Please describe what to research)'}
`
)

Step 2B: Breadth research — orchestrator-worker fan-out (default)

A single Sonnet pass under-covers a multi-part question. Anthropic's orchestrator-worker Research beat single-agent by 90.2% because each parallel worker spends a full, isolated context on one facet. The SKILL's main loop owns the orchestration (Masumi has no Task tool):

1. Decompose ${args} into 3–5 independent sub-questions, sized to complexity (a 2-way comparison → 2–3; a broad survey → 5).
2. Fan out one Masumi worker per sub-question, in PARALLEL:

   Task(subagent_type="team-shinchan:masumi", model="sonnet",
     prompt=`/team-shinchan:research worker. Parent question: ${args}
   Sub-question (yours only): ${subq}
   Use WebSearch + WebFetch. Return a tight cited brief: Key Findings (with source URLs) | Best Practices | Caveats. Flag low-confidence or conflicting claims explicitly.`)
   

3. Synthesize with one Masumi pass over the worker briefs:

   Task(subagent_type="team-shinchan:masumi", model="sonnet",
     prompt=`/team-shinchan:research synthesis. Original question: ${args}
   Worker briefs:
   ${worker_briefs}
   Produce ONE cited report — | Key Findings | Documentation | Best Practices | Caveats | — reconciling conflicts across workers and marking any single-source claim as low-confidence.`)
   

Launch workers in parallel (cap ~5, matching the 3–5 subagent guidance).

Step 2C: Claim-critical → offer native /deep-research (opt-in)

If the breadth research is claim-critical (security best practices, a hard-to-reverse vendor/architecture choice, anything feeding an irreversible Stage-2 decision), after Step 2B narrate a one-line opt-in: "This looks claim-critical — want me to run the native /deep-research for adversarial per-claim verification across more sources?" Do NOT auto-fire it; the user launches /deep-research. (Don't reimplement adversarial claim verification here — native deep-research already does it.)