Claude Powerpack

> Essential productivity tools for Claude Code: deep codebase analysis agent, expert consultation docs, code extraction, PR workflows, and more

 

Features

Code Researcher Agent

Deep codebase analysis specialist for answering complex questions about code:

• Execution Flow Tracing: Follows call chains across frontend/backend boundaries through every layer
• Usage Pattern Finding: Quantified counts with file:line references
• Architecture Mapping: Full hierarchies with ASCII diagrams
• Inconsistency Auditing: Severity-rated findings across all consumer layers
• Structural Search: ast-grep for AST-level pattern matching
• Git Investigation: History, blame, evolution tracking

Debrief Skill

End-of-session learning capture with progressive disclosure:

• Session Analysis: Scans conversation for mistakes, corrections, discoveries, and decisions
• Doc Audit: Checks CLAUDE.md, .claude/rules/, and project docs to avoid duplicates
• Progressive Disclosure: High-signal one-liners to CLAUDE.md, detailed knowledge to .claude/rules/
• Stale Entry Pruning: Identifies entries duplicated by docs, enforced by toolchain, outdated, or left as session debris
• Token Discipline: Enforces CLAUDE.md under 300 lines, routes domain knowledge to scoped rules
• User Approval: Presents all proposed changes before writing anything

Deep Interview Skill

Structured requirements gathering via interactive interviews:

• 7-Phase Process: Context → Scope → Approach → Priority → Details → Technical → Timeline
• GitHub Integration: Pulls issue context before asking questions
• PRD Generation: Auto-generates PRD with decision log and implementation checklist
• Best Practices: Multi-select questions, tradeoff presentation, stakeholder probing

Trim PR Skill

Pre-merge PR cleanup to trim accumulated complexity:

• Cruft Detection: Debug logging, commented code, unused imports, TODOs for done work
• Over-Engineering Check: Premature abstractions, dead code paths, speculative features
• Framework-Agnostic: Auto-detects Node.js, Python, Go, Rust for verification
• Risk-Rated Output: Change summary with Low/Medium/High risk assessment

Update PR Skill

Create comprehensive PR descriptions that document every meaningful change:

• Systematic Review: Inventories ALL changed files and commits
• Multi-Category: Features, bug fixes, tests, docs, config - nothing missed
• Structured Output: Organized sections for user impact and technical details
• GitHub Integration: Updates PR via gh pr edit

PR Audit Angles Skill

Multi-angle parallel audit of a large PR or branch before merge:

• Parallel Angles: One capable subagent per angle (backend, data/schema, API, frontend, auth, async, i18n, tests, completeness/parity), matched to the changed surface
• Repo-Anchored: Angles, hard rules, and commands are discovered from your repo's docs, configs, and CI — not hardcoded
• Lead, Not Verdict: Every candidate finding is verified against current code (with a file:line citation) before it reaches you
• Merge-Readiness Verdict: One synthesized report — Blockers / Important / New-concepts table / Verified-clean
• Pairs With Fix: Hand the verified findings straight to the PR Fix Angles skill

PR Fix Angles Skill

Parallel disjoint-lane fix campaign — the execution companion to PR Audit Angles:

• Disjoint Lanes: Findings partitioned by exclusive file ownership, so multiple subagents edit in parallel without merge conflicts
• Staged Hot Files: A lane whose findings share a file runs as sequential stages; independent lanes run concurrently
• Orchestrator-Owned Gates: The main agent runs the single final lint→build→test and never commits without you
• Test-First Fixes: The failing-test design is written into each fix prompt; assertions are never weakened to go green
• Skeptical Pre-Step: Optional find→adversarially-verify pass so agents never fix phantom or already-closed findings

Analyze Dependencies Skill

Changelog analysis for dependency updates with actionable recommendations:

• Multi-Mode: Preflight (before updating), post-update (after), or specific package research
• Ecosystem Support: npm and NuGet (auto-detected from project files)
• Multi-Source Research: Context7, GitHub Releases API, WebSearch, WebFetch
• Codebase Cross-Reference: Greps for actual usage of deprecated/changed APIs
• Structured Report: Breaking changes, new features, deprecations, security fixes, recommendations

Ask Expert Skill

Create comprehensive technical consultation documents for external expert review:

• Code Extraction: Full files, line ranges, or git diffs
• Size Tracking: Real-time progress with 125KB limit warnings
• Multi-file Support: Batch multiple files in single command
• Section Organization: Structured markdown headers
• Git Integration: Extract changes between branches/commits
• Config Files: Reusable extraction plans via JSON

Installation

Quick Install

1. Add the marketplace:

/plugin marketplace add propstreet/claude-powerpack

2. Install the plugin:

/plugin install claude-powerpack@propstreet

Or use the interactive menu:

/plugin

Then browse and select claude-powerpack from the available plugins.

After installation, restart Claude Code to activate the plugin.

Verify Installation

Check that the plugin is installed:

/plugin

You should see claude-powerpack in the list of installed plugins.

Usage

Code Researcher Agent

The code-researcher agent activates when Claude Code spawns a sub-agent for deep codebase analysis. It's used automatically by the Agent tool when research questions match its description. See Skills Included for example prompts.

Debrief Skill

Run at the end of a session to capture learnings. Claude reviews the conversation, identifies what was learned, audits existing docs, and proposes targeted updates. See Skills Included for example prompts.

Deep Interview Skill

Ask Claude to interview you about a feature before implementation. Claude will research context, ask structured questions, and generate a PRD. See Skills Included for example prompts.

Trim PR Skill

Ask Claude to trim your PR before merging. Claude will identify accumulated cruft, over-engineering, and suggest cleanups with risk ratings. See Skills Included for example prompts.

Update PR Skill

Ask Claude to update or prepare a PR description. Claude will systematically inventory all changes, categorize them, and create a structured summary. See Skills Included for example prompts.

PR Audit Angles Skill

Ask Claude to audit a PR or branch before merge. Claude splits the change into angles, runs one subagent per angle in parallel anchored to your repo's conventions, verifies each finding against current code, and returns a single merge-readiness verdict. See Skills Included for example prompts.

PR Fix Angles Skill

Ask Claude to fix all the verified findings. Claude partitions them into disjoint-file lanes, runs subagents in parallel without merge conflicts, owns the final lint→build→test gate, and leaves the changes for you to commit. See Skills Included for example prompts.

Analyze Dependencies Skill

Ask Claude to analyze dependency updates or run a preflight check before upgrading. Claude researches changelogs, cross-references your codebase, and generates a structured report. See Skills Included for example prompts.

Ask Expert Skill

Ask Claude to create consultation documents for external expert review. Claude will structure the problem, extract relevant code with size tracking, and format everything within the 125KB token limit. See Skills Included for example prompts.

Manual Code Extraction

You can also use the extraction script directly:

Basic extraction:

node scripts/extract-code.js \
  --track-size --output=doc.md \
  src/Component.vue tests/Component.test.ts

With sections:

node scripts/extract-code.js \
  --track-size --output=doc.md \
  --section="What Changed" \
  src/Service.cs:diff \
  --section="Implementation" \
  src/Service.cs src/Model.cs

Git diffs:

node scripts/extract-code.js \
  --track-size --output=doc.md \
  src/Service.cs:diff=master..feature-branch

Using config file:

node scripts/extract-code.js \
  --config=extraction-plan.json

See skills/ask-expert/EXAMPLES.md for complete usage examples.

Requirements

• Claude Code with plugin support
• Node.js 18+ (for code extraction script)
• Git (for diff functionality)
• gh CLI (for update-pr skill) - Install

Team Setup

Configure repository-level plugin installation by adding to your project's .claude/settings.json:

{
  "plugins": [
    "propstreet/claude-powerpack"
  ]
}

Team members who trust the repository folder will automatically have the plugin installed.

Documentation

Code Researcher Agent

• Agent Definition - Full agent prompt with methodology, tool guidance, and output calibration

Debrief Skill

• SKILL.md - Complete debrief workflow
• EXAMPLES.md - Good vs bad examples, routing decisions
• README - Quick reference

Deep Interview Skill

• SKILL.md - Complete interview workflow

Trim PR Skill

• SKILL.md - Complete cleanup workflow

Update PR Skill

• SKILL.md - Complete workflow for Claude
• EXAMPLES.md - Good vs bad PR examples
• README - Quick reference

PR Audit Angles Skill

• SKILL.md - Complete multi-angle audit workflow
• EXAMPLES.md - Structured schemas + parallel-orchestration script

PR Fix Angles Skill

• SKILL.md - Complete disjoint-lane fix-campaign workflow
• EXAMPLES.md - Fix schema + disjoint-lane orchestration script

Analyze Dependencies Skill

• SKILL.md - Complete analysis workflow
• README - Quick reference

Ask Expert Skill

• SKILL.md - Complete workflow for Claude
• EXAMPLES.md - Usage patterns and workflows
• README - Quick reference

Agents & Skills Included

code-researcher (v1.7.0)

Example prompts:

"Research how the auth flow works end-to-end"
"Find all error handling patterns and check for inconsistencies"
"Map the test infrastructure architecture"
"Trace the execution flow from the API endpoint to the database query"

What it does:

• Deep codebase analysis using ast-grep, ripgrep, git, and native Claude Code tools
• Traces execution flows across frontend/backend boundaries
• Finds usage patterns with quantified counts
• Audits for inconsistencies with severity-rated findings
• Maps architecture hierarchies with ASCII diagrams

Model guidance: Use a faster, lower-cost model for fast lookups and pattern counting; use your most capable model for deep architecture mapping and multi-layer flow tracing.

Allowed tools: Bash(ast-grep, rg, grep, find, git, jq, wc, head, tail, cat, sort, uniq, cut, tr, xargs, sed, awk, diff), Read, Glob, Grep

debrief (v1.5.0)

Example prompts:

"Debrief this session"
"Capture learnings from what we just did"
"Update project knowledge with what we learned"
"What did we learn? Save it."

What it does:

• Reviews session for mistakes, corrections, discoveries, and decisions
• Audits CLAUDE.md, .claude/rules/, and project docs for existing coverage
• Proposes targeted updates with progressive disclosure
• Enforces CLAUDE.md token discipline (under 300 lines, high-signal only)
• Gets user approval before writing any changes

Allowed tools: Bash(git), Read, Edit, Write, Glob, Grep, AskUserQuestion

deep-interview (v1.4.0)

Example prompts:

"Interview me about this feature before we start"
"Let's do a deep interview for issue #531"
"Help me gather requirements for the new authentication flow"
"Start a structured interview for this PRD"

What it does:

• Gathers context from GitHub issues and existing docs
• Asks structured questions through 7 phases
• Generates PRD with decision log and implementation checklist
• Probes for existing patterns and stakeholder knowledge

Allowed tools: Bash(git, gh), Read, Glob, Grep, AskUserQuestion, Write, Agent

trim-pr (introduced as simplify in v1.4.0, renamed in v1.8.0)

Example prompts:

"Trim this PR before merging"
"Clean up the accumulated cruft in this branch"
"Check for over-engineering before I merge"
"Review this PR for unnecessary complexity"

What it does:

• Identifies debug logging, commented code, unused imports
• Detects over-engineering and premature abstractions
• Auto-detects project type for verification commands
• Outputs risk-rated change summary table

Allowed tools: Bash(git, gh), Read, Edit, Glob, Grep

update-pr (v1.1.0)

Example prompts:

"Update the PR description"
"Prepare this PR for review"
"Document the changes in this branch"
"Write a comprehensive PR summary"

What it does:

• Inventories ALL changed files and commits systematically
• Categorizes changes: features, bug fixes, tests, docs, config
• Creates structured PR description with user impact section
• Saves to /tmp/pr-summary.md and updates PR via gh pr edit

Allowed tools: Bash, Read, Write, Edit, Glob, Grep

pr-audit-angles (v2.2.0)

Example prompts:

"Audit this PR before I merge"
"Review the branch from every angle"
"Am I introducing new concepts where patterns already exist?"
"Is this safe to merge?"

What it does:

• Splits a large diff into independent angles matched to the changed surface
• Dispatches one capable subagent per angle in parallel, anchored to the repo's own docs/conventions
• Forces uniform structured findings (Blockers / Important / New-concepts verdict / Verified-clean)
• Verifies every candidate against current code before reporting — a subagent claim is a lead, not a verdict
• Synthesizes one merge-readiness verdict; hands verified findings to pr-fix-angles

Allowed tools: Read, Grep, Glob, Bash(git, gh), Agent

pr-fix-angles (v2.2.0)

Example prompts:

"Fix all the audit findings, blocker to minor"
"Patch everything and make it merge-ready"
"Run a fix campaign on this PR"

What it does:

• Partitions verified findings into lanes of exclusive file ownership for conflict-free parallel edits
• Runs a shared-hot-file lane as sequential stages; independent lanes concurrently
• Writes the failing-test design into each fix prompt; never weakens assertions to go green
• Keeps the orchestrator in charge of the final lint→build→test gate and the commit decision
• Optional skeptical find→verify pre-step so agents never fix phantom or already-closed findings

Allowed tools: Read, Grep, Glob, Edit, Write, Bash(git, gh, npm, pnpm, yarn, npx, make, pytest, go, cargo, dotnet), Agent

analyze-deps (v1.9.0)

Example prompts:

"Analyze what changed after the dependency update"
"Run a preflight check before updating packages"
"What breaking changes are in vue 3.5 to 3.6?"
"Analyze deps for frontend only"

What it does:

• Detects npm and NuGet ecosystems from project files
• Preflight mode: checks what's outdated and researches before you update
• Post-update mode: detects changes via git diff and analyzes changelogs
• Fetches release notes from Context7, GitHub Releases, WebSearch
• Cross-references breaking changes against your actual codebase usage
• Generates structured report with recommendations

Allowed tools: Bash(git, ncu, npm, dotnet, gh, ls, find), Read, Glob, Grep, Agent, WebSearch, WebFetch, Context7

ask-expert (v1.0.0)

Example prompts:

"Create an expert consultation document for our authentication refactor"
"Prepare code for expert review about our API design"
"I need to ask an expert about our database schema"

What it does:

• Structures consultation documents with problem context
• Extracts code with real-time size tracking (125KB limit)
• Supports full files, line ranges, and git diffs
• Formats everything for external expert review

Allowed tools: Bash, Read, Write, Edit

Troubleshooting

Plugin not activating

1. Verify installation: /plugin 2. Restart Claude Code after installing 3. Try explicit activation phrases: "create an expert consultation document"

Script not found

# Verify script exists
ls scripts/extract-code.js

# Check Node.js version
node --version  # Should be 18+

File not found errors

# Check current directory
pwd

# Use absolute paths for files in your repo
node scripts/extract-code.js \
  /absolute/path/to/file.cs

See skills/ask-expert/README.md for complete troubleshooting guide.

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions

Contributing

Contributions welcome! Please see CONTRIBUTING.md for guidelines.

License

MIT - See LICENSE for details.

Credits

Created by Propstreet for the Claude Code community.

Inspired by the need to effectively communicate complex technical context to external experts and AI consultants while respecting token limits.