arndvs/explore

Explore

Output "Read Explore skill." to chat to acknowledge you read this file.

When tasked with understanding, auditing, or investigating any part of a codebase, spawn multiple sub-agents using the explore verb to maximize file coverage and depth. Use the search_subagent tool with explore in the prompt to trigger deep traversal mode, allowing the agent to recursively search through directories and files for relevant information.

When to Use This

Trigger this approach when asked to:

• Understand how a feature works
• Audit or review a codebase
• Investigate a bug or behavior
• Map out data flow or architecture
• Prepare before building a new feature

How to Execute

0. HUD event — emit a start event so the HUD tracks this exploration:

   source ~/dotfiles/bin/write-hud-state.sh
   write_hud_event "info" "explore: started — $TOPIC"
   

   Emit again at synthesis (explore: completed — N sub-agents, M files covered).
1. Decompose the topic into distinct areas of concern
2. Spawn a dedicated sub-agent for each area using the explore keyword in the sub-agent prompt
3. Synthesize the results from all sub-agents into a single cohesive summary

Example decomposition for "How does authentication work?":

• Explore how sessions are created and stored
• Explore how middleware protects routes
• Explore how tokens are issued and validated
• Explore how logout and expiry are handled

Rules

• Always use the search_subagent tool and the word explore when spawning sub-agents — this triggers deep file traversal mode to run MULTIPLE sub-agents in PARALLEL, maximizing coverage and depth
• Each sub-agent should have a narrow, specific focus — avoid broad prompts per agent
• Do not rely on a single agent to cover the full topic
• After all sub-agents complete, produce a unified summary with cross-references between areas
• If a sub-agent surfaces something unexpected, spawn an additional agent to follow up

Why This Matters

A single agent pass typically reads 6–10 files. A dedicated explore sub-agent will aggressively search the repo, calling 20–30+ tools and covering far more ground. For any non-trivial exploration task, a single pass will produce an incomplete picture.