Skip to content

jw409/kinderpowers

Repository files navigation

kinder·powers

Make your AI agent actually finish what it starts.

Kinderpowers is a plugin for Claude Code that teaches agents engineering judgment — verify before claiming "done", branch when uncertain, run tests before committing, coordinate parallel workers without merge conflicts.

claude plugin marketplace add jw409/kinderpowers
claude plugin install kinderpowers

The Progression Model

Most agent frameworks are flat. Kinderpowers grows with the agent.

graph LR
    L1["<b>L1: Coding Assistant</b><br/>6 skills · 1 agent<br/><i>TDD · debug · verify</i>"]
    L2["<b>L2: Agentic Worker</b><br/>17 skills · beads tracking<br/><i>plan · execute · remember</i>"]
    L3["<b>L3: Team Orchestrator</b><br/>3 skills · 14 agents<br/><i>decompose · dispatch · integrate</i>"]
    L4["<b>L4: Dark Factory</b><br/>42 commands · 3 rules<br/><i>roadmap → plan → execute → ship</i>"]
    L1 --> L2 --> L3 --> L4
    style L1 fill:#e8f5e9,stroke:#4caf50
    style L2 fill:#e3f2fd,stroke:#2196f3
    style L3 fill:#fff3e0,stroke:#ff9800
    style L4 fill:#fce4ec,stroke:#e91e63
Loading

L1 — Your agent stops saying "looks correct" and starts showing you test output. Six skills auto-inject: brainstorming, TDD, systematic debugging, verification, code review (giving and receiving).

L2 — Your agent plans before coding, remembers across sessions, and works in git worktrees so it never blocks your branch. Seventeen skills covering strategic planning, metathinking, research extraction, beads-based tracking, and more.

L3 — Five modules to update? Spawn five workers with non-overlapping file domains. Twenty minutes, zero conflicts. Includes multi-perspective review that selects lenses (security, edge cases, resilience) based on what could actually break.

L4 — Full lifecycle automation. You describe a feature; the agent roadmaps it, plans phases, executes with atomic commits, verifies against requirements, and ships a PR. You review at phase boundaries. Forty-two GSD commands, three hookify rules that ship disabled until you opt in.


Before / After

Without kinderpowers:

You:   "Fix the date parsing bug"
Agent: "I've updated line 47 to use Date.parse()."
You:   "Did you run the tests?"
Agent: "The logic looks correct."
You:   *runs tests* — 3 failures, the "fix" broke timezone handling

With kinderpowers:

You:   "Fix the date parsing bug"
Agent: "Reproduced: test_parse_iso_date fails with '2024-03-15T10:00:00Z'
        Root cause: Date.parse() returns local time, not UTC
        Fix: new Date(str).toISOString() preserves timezone
        Regression test added: test_timezone_preservation
        Full suite: 24/24 pass

        $ npm test -- --grep 'date'
        PASS  ✓ test_parse_iso_date (3ms)
             ✓ test_timezone_preservation (2ms)  ← new"

The agent didn't get smarter. It got disciplined. Skills auto-inject based on what you're doing — touch a test file and TDD activates, claim "done" and verification asks for evidence.


MCP Servers

Two Rust-native MCP servers, shipped with pre-built binaries (linux-x86_64, macOS-arm64). No Rust toolchain required.

kp-github — 7x fewer tokens than the official plugin

The official Claude Code GitHub plugin returns raw API responses. Listing 5 issues burns ~7,700 tokens on avatar URLs, node IDs, and empty arrays. kp-github runs a 5-stage compression pipeline and returns ~1,100 tokens for the same query.

63 tools. Full superset of the official plugin plus Actions, Labels, Compare, and Release Create. Every read tool accepts fields and format parameters.

484 unit tests, 23 integration tests, 97% coverage. 7.3MB binary.

kp-sequential-thinking — structured reasoning with hints

Branching, confidence tracking (with Dunning-Kruger detection), abstraction layers, exploration mode, branch merging, per-model profiles (Claude, Gemini, DeepSeek, Grok, Llama/Nemotron), and JSONL logging for learning pipelines.

Six hint types surface observations about reasoning patterns. The agent decides what to act on.

106 tests, 95.6% coverage. 4.6MB binary.

# Install both
cd mcp-servers && ./install.sh

Who Is This For

Solo developers using Claude Code who are tired of agents that guess instead of verify.

Team leads coordinating parallel agent workers on shared codebases.

Store owners building on Unified Commerce Platform (UCP) — kinderpowers + UCP gives your agent the discipline to handle payment flows, inventory sync, and multi-channel operations without cutting corners.

Anyone building with agents who wants a progression model instead of a flat skill dump.


What's Inside

Everything auto-routes. You don't need to memorize this.

Layer Count Details
Skills 27 Auto-injected when relevant files are touched (manifest)
Agents 14 6 kinderpowers + 8 GSD, spawned by skills and commands (agents)
GSD Commands 42 Full lifecycle — or just /gsd:quick and /gsd:autonomous
Hookify Rules 3 Ship disabled. Verification-required, discovery-before-creation, brainstorm-before-build
MCP Servers 2 kp-github (63 tools), kp-sequential-thinking

The philosophy: skills are invitations, not commands. Every recommendation documents the cost of skipping it, so agents make informed trade-offs instead of blindly obeying.

Enforcement Rules

All rules ship disabled. You opt in.

  • verification-required — blocks "done" claims without evidence (test output, curl responses, screenshots)
  • discovery-before-creation — warns before creating files without searching for existing solutions
  • brainstorm-before-build — warns before writing 100+ lines without design discussion

Scanner

scanner.py detects compulsion language in skill files — directive words that remove agent judgment. Five severity tiers, CI integration via --check.

uv run python scanner.py --check skills/

Installation

claude plugin marketplace add jw409/kinderpowers
claude plugin install kinderpowers

For manual installation or other platforms (Cursor, Codex, OpenCode):

git clone https://github.com/jw409/kinderpowers.git ~/.claude/plugins/kinderpowers
cd ~/.claude/plugins/kinderpowers && ./setup.sh

setup.sh creates the GSD runtime symlink, installs hookify rules if present, and sets up the agent-outcome-logger hook. Idempotent. GSD commands and agents register automatically via the plugin system.


Credits

  • superpowers by Jesse Vincent — craft philosophy, skill format, scanner, hook system
  • get-shit-done by David Braun — lifecycle engine, 42 commands, 8 parameterized agents
  • hookify by Diego Perez — enforcement rule format, Claude Code hook framework
  • jw409 — progression model, agency-preserving philosophy, council mode, MCP servers

License

MIT — see LICENSE.


claude plugin marketplace add jw409/kinderpowers
claude plugin install kinderpowers

About

Agency-preserving workflow skills for AI agents

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors