Explore

• Use explore when the solution space is open: user knows the problem, not the options
• Use /empire-research:compare instead when user already has a known set of options to evaluate head-to-head
• If user input names specific options (A vs B vs C), suggest /empire-research:compare and confirm before proceeding here

• Read conversation for problem statement, scope, constraints, success criteria
• Signals to read:
• Explicit user description of the problem
• Recent code or files providing technical context
• Stated constraints (budget, timeline, stack, team size)
• Prior approaches already ruled out
• Definition of "good enough" outcome
• If problem statement unclear → ask one clarifying question at a time
• If structured choices help → use AskUserQuestion with concrete options
• MUST state inferred problem statement back to user before any dispatch
• MUST get user confirmation on problem statement
• MUST NOT dispatch any agent until problem is confirmed

• After problem confirmed, dispatch ONE research agent for broad enumeration
• Agent names vary by environment; do not assume a specific agent exists
• Inspect available subagents via the Agent tool's subagent_type parameter
• Pick the available agent whose name/description best matches general research synthesis or broad information retrieval; if multiple candidates fit, prefer the most specific; if none fit, use the most general research-oriented agent available
• Shallow agent instructions:

• Enumerate 3–5 candidate approaches only
• One short paragraph per approach — no deep evaluation
• Required output format:

```

... ```

• Cap response under 300 words

• Present shallow-scan output to user verbatim before proceeding

• Gate exists because deep-dive spawns one parallel agent per approach (real cost) — user steers spend toward the approaches worth researching
• After shallow scan, present results and ask user:
• Which approaches to deep-dive (may pick multiple)
• Whether to add, remove, or reframe any approach
• MUST wait for explicit user selection before deep dispatch
• MUST NOT infer selection and proceed silently
• If user requests a different approach not in list → add it, confirm updated list

• After approaches selected, dispatch the deep research one of two ways
• Preferred — Workflow tool available:

• Invoke the bundled deep-dive workflow; it fans out one researcher per approach with structured pros/cons/fit:

`` Workflow({ scriptPath: "${CLAUDE_PLUGIN_ROOT}/workflows/explore-deepdive.js", args: { problem, constraints, successCriteria, approaches: [{ name, description }] }, }) ``

• Surface the workflow's log() lines as progress
• Feed the returned approaches[] into consolidated-report
• Skip agent-selection and parallel-deep-dispatch — the workflow owns dispatch

• Fallback — Workflow tool unavailable: use agent-selection then parallel-deep-dispatch below

• Fallback path — only when the Workflow tool is unavailable (see dispatch-mode)
• Pick one deep agent per selected approach
• Agent names vary by environment; do not assume a specific agent exists
• Inspect available subagents via the Agent tool's subagent_type parameter
• For each selected approach, identify its dominant signal from these categories:
• General synthesis, multi-source aggregation
• Fast targeted retrieval, known-solution space
• Quantitative datasets, benchmarks, numerical evidence
• Peer-reviewed or scientific evidence
• Emerging-tech trajectory, trend analysis
• For each signal that applies, pick the available agent whose name/description best matches; if multiple candidates fit, prefer the most specific; if none fit, use the most general research-synthesis agent available
• MUST always include at least one general research-synthesis agent to anchor the roster
• List chosen agent per approach (using its actual subagent_type value) + one-line rationale BEFORE dispatch
• If confident in every pick → dispatch immediately
• If uncertain about any pick → confirm roster with user before dispatch; allow swaps

• Send single message with multiple Agent tool calls (one per approach)
• Each agent receives:
• Original confirmed problem statement
• The specific approach assigned to them
• All known constraints and success criteria
• Output format instruction (see below)
• "Do NOT post findings to any external system. Report in chat only."
• Required deep agent output format:

``` Approach:

Summary:

Pros:

Cons:

Key Evidence / Citations:

Fit Rating: — ```

• Cap each agent response under 500 words

• After all deep agents return, produce consolidated report
• If the workflow returns `stats.researched

• MUST gather and confirm problem context before any agent dispatch
• MUST clarify ambiguity before shallow scan
• MUST confirm shallow results with user before deep dispatch
• MUST dispatch deep research via the explore-deepdive workflow when the Workflow tool is available; else dispatch deep agents in parallel (single message, multiple tool uses)
• MUST keep all findings local in chat only
• MUST NOT post to Slack, GitHub, Jira, or any external system unless user explicitly authorizes
• MUST NOT implement chosen approach — recommendation only
• MUST NOT proceed through any gate without explicit user confirmation
• If zero suitable research-synthesis agents exist in environment → MUST stop and tell user; never inline-impersonate a researcher

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: marcoskichel
• Source: marcoskichel/empire
• License: MIT

Install and usage instructions live in the source repository linked above.