MCPInstall

explore

Meta-skill for internal codebase exploration at varying depths (quick/deep/architecture)

Installs
577
Last 24h
—
Trending rank
—
Security
Not audited

Install

Works with every agent the skills CLI supports.

npx skills add https://github.com/parcadei/continuous-claude-v3 --skill explore
  1. 01

    Install

    Run the command in your project. The CLI asks which agents to install for.

  2. 02

    Describe the task

    Ask your agent for something this skill covers. It loads explore automatically when relevant.

  3. 03

    Review the result

    The agent follows the SKILL.md steps, scripts and references bundled with the skill.

SKILL.md

The instructions your agent reads when the skill is triggered.

Explore - Internal Codebase Exploration

Meta-skill for exploring an internal codebase at varying depths. READ-ONLY workflow - no code changes.

Usage

/explore <depth> [options]

Question Flow (No Arguments)

If the user types just /explore with no or partial arguments, guide them through this question flow. Use AskUserQuestion for each phase.

Phase 0: Workflow Selection

question: "How would you like to explore?"
header: "Explore"
options:
  - label: "Help me choose (Recommended)"
    description: "I'll ask questions to pick the right exploration depth"
  - label: "Quick - fast overview"
    description: "Chain: tldr tree → tldr structure (~1 min)"
  - label: "Deep - comprehensive analysis"
    description: "Chain: onboard → tldr → research → document (~5 min)"
  - label: "Architecture - layers & dependencies"
    description: "Chain: tldr arch → call graph → layer mapping (~3 min)"

Mapping:

  • "Help me choose" → Continue to Phase 1-4 questions
  • "Quick" → Set depth=quick, skip to Phase 2 (scope)
  • "Deep" → Set depth=deep, skip to Phase 2 (scope)
  • "Architecture" → Set depth=architecture, skip to Phase 2 (scope)

If Answer is Unclear (via "Other"):

question: "I want to understand how deep you want to explore. Did you mean..."
header: "Clarify"
options:
  - label: "Help me choose"
    description: "Not sure - guide me through questions"
  - label: "Quick - fast overview"
    description: "Just want to see what's here"
  - label: "Deep - comprehensive analysis"
    description: "Need thorough understanding"
  - label: "Neither - let me explain differently"
    description: "I'll describe what I need"

Phase 1: Exploration Goal

question: "What are you trying to understand?"
header: "Goal"
options:
  - label: "Get oriented in the codebase"
    description: "Quick overview of structure"
  - label: "Understand how something works"
    description: "Deep dive into specific area"
  - label: "Map the architecture"
    description: "Layers, dependencies, patterns"
  - label: "Find where something is"
    description: "Locate specific code/functionality"

Mapping:

  • "Get oriented" → quick depth
  • "Understand how" → deep depth
  • "Map architecture" → architecture depth
  • "Find where" → quick with --focus

Phase 2: Scope

question: "What area should I focus on?"
header: "Focus"
options:
  - label: "Entire codebase"
    description: "Explore everything"
  - label: "Specific directory or module"
    description: "I'll specify the path"
  - label: "Specific concept/feature"
    description: "e.g., 'authentication', 'API routes'"

If "Specific directory" or "Specific concept" → ask follow-up for the path/keyword.

Phase 3: Output Format

question: "What should I produce?"
header: "Output"
options:
  - label: "Just tell me what you find"
    description: "Interactive summary in chat"
  - label: "Create a documentation file"
    description: "Write to thoughts/shared/docs/"
  - label: "Create handoff for implementation"
    description: "Prepare context for coding agent"

Mapping:

  • "Documentation file" → --output doc
  • "Handoff for implementation" → --output handoff

Phase 4: Entry Point (Architecture only)

If architecture depth selected:

question: "Where should I start the analysis?"
header: "Entry point"
options:
  - label: "Auto-detect (main, cli, app)"
    description: "Find common entry points"
  - label: "Specific function/file"
    description: "I'll specify the entry point"

Summary Before Execution

Based on your answers, I'll run:

**Depth:** deep
**Focus:** "authentication"
**Output:** handoff
**Path:** src/

Proceed? [Yes / Adjust settings]

Depths

DepthTimeWhat it does
quick~1 mintldr-explorer only - fast structure overview
deep~5 minonboard + tldr-explorer + research-codebase + write doc
architecture~3 mintldr arch + call graph + layer mapping + circular dep detection

Options

OptionDescriptionExample
--focus "area"Focus on specific area--focus "auth", --focus "api"
--output handoffCreate handoff for next agent--output handoff
--output docCreate documentation file--output doc
--entry "func"Start from specific entry point--entry "main", --entry "process_request"

Examples

# Quick structure overview
/explore quick

# Deep exploration focused on auth
/explore deep --focus "auth" --output doc

# Architecture analysis from specific entry
/explore architecture --entry "cli" --output handoff

# Quick focused exploration
/explore quick --focus "hooks"

Workflow Details

Quick Depth

Fast structure overview using tldr-explorer. Best for:

  • Initial orientation
  • Quick questions about structure
  • Finding where things are

Steps:

  1. Run tldr tree for file structure
  2. Run tldr structure for codemaps
  3. If --focus provided, run tldr search for targeted results
  4. Return summary

Commands:

# 1. File tree
tldr tree ${PATH:-src/} --ext .py

# 2. Code structure
tldr structure ${PATH:-src/} --lang python

# 3. Focused search (if --focus provided)
tldr search "${FOCUS}" ${PATH:-src/}

Deep Depth

Comprehensive exploration with documentation output. Best for:

  • First time in a codebase
  • Preparing for major work
  • Creating reference documentation

Steps:

  1. Check if onboarded (look for .claude/cache/tldr/), if not run onboard
  2. Run tldr-explorer for structure
  3. Spawn research-codebase agent for patterns
  4. Write findings to doc or handoff

Subprocess:

# 1. Onboard check
if [ ! -f .claude/cache/tldr/arch.json ]; then
    # Spawn onboard agent
fi

# 2. Structure analysis
tldr structure src/ --lang python
tldr calls src/

# 3. Research patterns (via scout agent)
Task: research-codebase → "Document existing patterns in ${FOCUS:-codebase}"

# 4. Write output
→ thoughts/shared/research/YYYY-MM-DD-explore-{focus}.md
→ OR thoughts/shared/handoffs/{session}/explore-{focus}.yaml

Architecture Depth

Architecture-focused analysis with layer detection. Best for:

  • Understanding system boundaries
  • Preparing for refactoring
  • Identifying coupling issues

Steps:

  1. Run tldr arch for layer detection
  2. Run tldr calls for cross-file call graph
  3. Analyze entry/middle/leaf layers
  4. Detect circular dependencies
  5. Map architectural boundaries

Commands:

# 1. Architecture detection
tldr arch ${PATH:-src/}
# Returns: entry_layer, middle_layer, leaf_layer, circular_deps

# 2. Call graph
tldr calls ${PATH:-src/}
# Returns: edges, nodes

# 3. Impact analysis from entry point (if --entry provided)
tldr impact ${ENTRY} ${PATH:-src/} --depth 3

Output Structure:

layers:
  entry: [routes.py, cli.py, main.py]  # Controllers/handlers
  middle: [services.py, auth.py]        # Business logic
  leaf: [utils.py, helpers.py]          # Utilities

call_graph:
  total_edges: 142
  hot_paths: [process_request → validate → authorize]

circular_deps:
  - [module_a, module_b]  # A imports B, B imports A

boundaries:
  - name: API layer
    files: [src/api/*]
    calls_to: [src/services/*]

Output Formats

--output doc

Creates: thoughts/shared/research/YYYY-MM-DD-explore-{focus}.md

---
date: {ISO timestamp}
type: exploration
depth: {quick|deep|architecture}
focus: {focus area or "full"}
commit: {git hash}
---

# Codebase Exploration: {focus}

## Summary
{High-level findings}

## Structure
{File tree / codemaps}

## Architecture
{Layer analysis - for architecture depth}

## Key Components
{Important files and their roles}

## Patterns Found
{Existing patterns - for deep depth}

## References
- `path/to/file.py:line` - Description

--output handoff

Creates: thoughts/shared/handoffs/{session}/explore-{focus}.yaml

---
type: exploration
ts: {ISO timestamp}
depth: {quick|deep|architecture}
focus: {focus area}
commit: {git hash}
---

summary: {One-line summary of findings}

structure:
  entry_points: [{main.py}, {cli.py}]
  key_modules: [{auth.py}, {routes.py}]
  test_coverage: [{tests/}]

architecture:
  layers:
    entry: [{files}]
    middle: [{files}]
    leaf: [{files}]
  circular_deps: [{pairs}]

findings:
  - {key finding with file:line}

next_steps:
  - {Recommended action based on exploration}

refs:
  - path: {file.py}
    role: {what it does}

Integration with /build

The explore skill is designed to feed into /build brownfield:

# Step 1: Explore to understand
/explore architecture --output handoff

# Step 2: Build with context from exploration
/build brownfield --from-handoff thoughts/shared/handoffs/session/explore-full.yaml

Implementation

When user invokes /explore <depth> [options]:

Parse Arguments

depth = args[0]  # quick | deep | architecture
focus = extract_option(args, "--focus")
output = extract_option(args, "--output")  # handoff | doc
entry = extract_option(args, "--entry")

Execute Based on Depth

Quick:

# Just tldr commands, no agents
tldr tree ${src_dir} --ext .py
tldr structure ${src_dir} --lang python
if [ -n "$focus" ]; then
    tldr search "$focus" ${src_dir}
fi

Deep:

# 1. Check/run onboard
if [ ! -f .claude/cache/tldr/meta.json ]; then
    # Spawn onboard agent via Task tool
fi

# 2. Structure
tldr structure src/ --lang python

# 3. Research (spawn scout agent)
# Task tool with subagent_type: "scout"
# Prompt: "Research patterns in ${focus:-codebase}"

# 4. Write output
# → doc or handoff based on --output

Architecture:

# 1. Arch detection
arch_output=$(tldr arch ${src_dir})

# 2. Call graph
calls_output=$(tldr calls ${src_dir})

# 3. Impact from entry (if provided)
if [ -n "$entry" ]; then
    impact_output=$(tldr impact $entry ${src_dir} --depth 3)
fi

# 4. Synthesize and write output

Key Principles

  1. READ-ONLY - This skill never modifies code
  2. Uses scout, not Explore - Per project rules, scout (Sonnet) over Explore (Haiku)
  3. Token-efficient - Uses tldr commands (95% savings over raw reads)
  4. Outputs to shared locations - thoughts/shared/research/ or handoff directory
  5. Entry point to /build - Exploration handoffs feed into brownfield builds

Related Skills

SkillWhen to Use
tldr-explorerDirect tldr commands (used internally by explore)
tldr-codeSpecific analysis commands (cfg, dfg, slice)
onboardFirst-time project setup (used by deep depth)
research-codebasePattern documentation (used by deep depth)
create_handoffHandoff format (used by --output handoff)

Troubleshooting

tldr not found:

# Check if installed
which tldr
# Install if missing
uv tool install llm-tldr
# or: pip install llm-tldr

No Python files found:

# Check language, adjust --lang
tldr structure src/ --lang typescript  # or go, rust

Empty architecture output:

# May need to specify src directory
tldr arch ./  # Current directory
tldr arch src/  # Explicit src

FAQ

What does the explore skill do?

Meta-skill for internal codebase exploration at varying depths (quick/deep/architecture)

How do I install explore?

Run npx skills add https://github.com/parcadei/continuous-claude-v3 --skill explore in your project. The skills CLI detects the agents you use and copies the skill into the right folder. For Claude Code you can also place the skill folder at ~/.claude/skills/explore/ (personal) or .claude/skills/explore/ (project).

Which AI agents can use explore?

Any agent that supports the open Agent Skills format (a folder with a SKILL.md file), including Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI, Windsurf, Cline, OpenCode and more.

Is explore safe to install?

No partner security audit is available yet. Skills can run commands and scripts, so read the SKILL.md and any bundled files before installing.

Who maintains explore?

It is published by parcadei in parcadei/continuous-claude-v3. It has 577 recorded installs.

More from parcadei/continuous-claude-v3

View all →

discovery-interview

parcadei/continuous-claude-v3

Deep interview process to transform vague ideas into detailed specs. Works for technical and non-technical users.

4.1K installs

premortem

parcadei/continuous-claude-v3

Identify failure modes before they occur using structured risk analysis

886 installs

math

parcadei/continuous-claude-v3

Unified math capabilities - computation, solving, and explanation. I route to the right tool.

github-search

parcadei/continuous-claude-v3

Search GitHub code, repositories, issues, and PRs via MCP

AI & Agents732 installs

dead-code

parcadei/continuous-claude-v3

Find unused functions and dead code in the codebase

refactor

parcadei/continuous-claude-v3

Code refactoring workflow - analyze → plan → implement → review → validate

Popular code quality & workflow skills

View category →

grill-me

mattpocock/skills

A relentless interview to sharpen a plan or design.

Scan a codebase for deepening opportunities, present them as a visual HTML report, then grill through whichever one you pick.

handoff

mattpocock/skills

Compact the current conversation into a handoff document for another agent to pick up.

domain-modeling

mattpocock/skills

Build and sharpen a project's domain model. Use when discussing codebase terminology, writing or editing a GLOSSARY.md, or recording or editing an ADR.

codebase-design

mattpocock/skills

Shared vocabulary for designing deep modules. Use when the user wants to design or improve a module's interface, find deepening opportunities, decide where a seam goes, make code more testable or AI-navigable, or when another skill needs the deep-module vocabulary.

diagnosing-bugs

mattpocock/skills

Diagnosis loop for hard bugs and performance regressions. Use when the user says "diagnose"/"debug this", or reports something broken/throwing/failing/slow.