jk-skills

A heavyweight Claude Code plugin for planning, executing, and shipping software. It replaces several plugins with an integrated system: research → interview → design → adversarial review → implementation plan → parallel execution → verification.

This is opinionated. It expands scope aggressively, demands TDD, runs adversarial review panels, and won't let you ship without proving the work is done. If you want something lighter, this isn't it.

What It Does

Planning — Before writing code, jk-plan runs deep codebase research with parallel explorer agents, interviews you to understand the problem, has multiple architects independently design solutions, then runs a 6-reviewer adversarial panel to tear the design apart. It cycles until no critical issues remain.

Execution — jk-execute takes the plan and runs it in one of four modes:

Mode	How it works	Best for
Deep	One orchestrator dispatches subagents per task, reviews between tasks	Most work — tightly coupled tasks, refactoring
Direct	Main thread does the work, you see everything	Risky or uncertain work where you want full visibility
Swarm	Parallel subagents on independent files simultaneously	Bulk changes, 3+ independent tasks
Care	Like Deep but pauses at meaningful checkpoints for your review	High-stakes changes, unfamiliar codebases

Verification — jk-prove-it mechanically verifies the work before you ship. Runs tests, checks the diff, generates a ship report. No "I think it works" — evidence or it didn't happen.

Knowledge — jk-remember persists what was learned to the right place: CLAUDE.md for project conventions, docs/ for deeper knowledge, auto memory for your preferences. Runs at the end of planning and execution so knowledge compounds across sessions.

Reflection — jk-reflect steps back and challenges the current direction. Gut check first, then structured analysis. Can dispatch a fresh subagent for an unbiased perspective on complex decisions.

Burn Rate — jk-burn-rate lets you control token spending: max (opus everything, go wide), standard (balanced), or light (efficient, cheaper models where quality won't suffer). Never weakens core discipline — just controls how aggressively to spend on discretionary work.

Philosophy

Code is free. Expand scope relentlessly. Refactor always. Ask more questions. Build on what's already known. Run autonomously — blocking must be conscious and explicit. Every task is an opportunity to leave the codebase better.

Full text: invoke /jk-philosophy in a session.

Installation

Claude Code Marketplace

/plugin marketplace add JeremyKennedy/jk-skills
/plugin install jk-skills@jk-skills

Nix Flake (NixOS / home-manager)

# flake.nix inputs:
jk-skills = {
  url = "git+https://git.jeremyk.net/jeremy/jk-skills.git";
  inputs.nixpkgs.follows = "nixpkgs";
};

# home-manager config:
imports = [ inputs.jk-skills.nixosModules.default ];
programs.jk-skills.enable = true;

Pick one, not both. Using both creates duplicate skills.

Supersedes

If you have any of these installed, uninstall them — jk-skills absorbs and improves on their functionality:

• superpowers — all skills + code-reviewer agent
• feature-dev — code-explorer and code-architect agents
• pr-review-toolkit — silent-failure-hunter, test-analyzer, doc-analyzer agents
• code-review — jk-code-review covers this
• code-simplifier — overlaps with built-in review
• claude-md-management — jk-remember handles CLAUDE.md + docs/ + memory routing

Run /jk-plugin-check to detect conflicts automatically.

All Skills

Core Workflow (roughly in order of use)

Skill	What it does
jk-brainstorm	Lightweight conversational ideation with optional visual browser companion
jk-plan	Research → interview → design → review panel → implementation plan
jk-execute	Run plans in Deep / Direct / Swarm / Care mode with TDD and code review
jk-prove-it	Mechanical verification, self-review, ship report
jk-finish-branch	Analyze branch state, present merge/push/PR/cleanup options
jk-remember	Smart knowledge routing: CLAUDE.md / docs/ / auto memory. Scales from quick save to full doc audit

Support Skills

Skill	What it does
jk-reflect	Step back, challenge assumptions, optionally dispatch fresh subagent for outside perspective
jk-burn-rate	Session-level token spending control: max / standard / light

Development Discipline

Skill	What it does
systematic-debugging	Root cause investigation before proposing fixes
test-driven-development	TDD: failing test → implement → refactor
verification-before-completion	Evidence before claims — run verification, confirm output
jk-code-review	Dispatch code-reviewer agent (automatic in jk-execute, manual for ad-hoc)
jk-receive-review	Handle code review feedback with rigor — challenge assumptions, verify independently

Coordination

Skill	What it does
dispatching-parallel-agents	Parallel subagent coordination with context isolation
using-git-worktrees	Isolated workspace setup with safety verification
jk-converse	Structured async conversation between two agents via shared markdown file

Meta

Skill	What it does
using-jk-skills	Auto-loaded via SessionStart hook. Skill discovery and routing.
jk-philosophy	Foundational philosophy: code is free, expand scope, refactor always, heavyweight autonomy
jk-plugin-check	Detect installed plugins that jk-skills supersedes
writing-skills	Skill authoring and verification

Agents

Agent	Used by	What it does
code-reviewer	jk-execute, jk-code-review	Review code against plan and standards
code-explorer	jk-plan Phase 1	Deep codebase analysis
code-architect	jk-plan Phase 4	Design feature architectures
silent-failure-hunter	jk-execute round table	Audit error handling for silent failures
test-analyzer	jk-execute round table	Behavioral test coverage analysis
doc-analyzer	jk-execute round table	Documentation accuracy and staleness detection

License

GPL v3. See LICENSE.

Derived from superpowers (MIT), claude-plugins-official (Apache 2.0). See ATTRIBUTION.md.