v2.0.0 — Living coherence + self-improving skills

The context system
Claude Code deserves.

Persistent memory. Reusable slash commands. Hooks automation. Skills that learn from their own failures. A production-ready framework so your AI stops starting from zero every session.

Get started → View on GitHub
Why ClaudeKit

Claude is powerful.
But every session starts blank.

ClaudeKit turns Claude Code from a brilliant amnesiac into an assistant that remembers your patterns, avoids its recurring mistakes, and actually gets better over time.

Lost context between sessions
Persistent memory system with tiered storage
Repeating the same instructions
Reusable slash commands: /focus /investigate
No workflow automation
Hooks for auto-formatting, validation, security gates
Scattered project knowledge
Structured patterns and quick-reference docs
Recurring mistakes
Self-improving skills that learn from errors
What's inside

Four systems. One kit.

Each pillar stands on its own — drop in the memory system and skip hooks if you want. But they're designed to work together, and they really shine when they do.

🧠

Memory

Tiered storage — quick-reference for hot patterns, structured patterns by domain, sprint-level episodic notes. Claude checks it first, every time.

⌘

Commands

10 slash commands for focusing, investigating, sprint planning, and bootstrapping projects. Each one loads the right context before it acts.

🔗

Hooks

Python hooks for security gates, error detection, skill suggestions, and auto-formatting. Runs before, during, and after tool calls.

✦

Skills

Auto-invoked capabilities, plus a self-improvement loop where Claude catalogs its own errors and turns them into working skills.

Slash commands

Workflows, not one-liners.

Every command loads memory first, acts second. You get consistent behavior across sessions without repeating yourself.

/focus [task]
Load context and concentrate on a specific task. Pulls quick-reference, pain points, and related files.
/investigate [question]
Search and analyze the codebase to answer a specific question with grounded evidence.
/deep-investigate [problem]
Multi-agent parallel investigation for complex problems. Dispatches sub-agents and synthesizes findings.
/brainstorm-design [element]
Generate design concepts with external research. Web search + pattern matching.
/plan-as-group [problem]
Three technical experts collaboratively plan a solution. Surfaces trade-offs and dissent.
/sprint-plan [week]
Automate weekly sprint planning. Synthesizes memory, pain points, and open tasks into a plan.
/orchestrate-tasks
Launch parallel task execution across independent work streams.
/bootstrap-project [dir]
Full codebase analysis and context setup. One command to onboard Claude to any repo.
/coherence
Run one cycle of the coherence skill — keeps a living repo understanding via the document surface.
/hooks-analyzer
Analyze your CLAUDE.md for opportunities to automate with hooks.
/update-template
Check for and apply ClaudeKit template updates. Preserves your memory and pain points.
Skills

Capabilities that activate themselves.

Skills auto-invoke based on conversation context — no commands to remember. They also include something most kits don't: a feedback loop where Claude's own mistakes become new skills.

User-invocable

trigger via natural language
project-builder
Break down projects into prioritized task lists. Triggers on "build" or "implement" requests.
pain-point-manager
Capture, update, and archive development friction. Triggers on mentions of blockers or workarounds.
investigation-analysis
ROI analysis for feature requests and refactors. Triggers on "should we build this?"
bloat-manager
Prevents unbounded growth of system artifacts. Runs automatically on Mondays.
coherence
Maintains a living, self-updating understanding of a repo via a decomposable document surface.

Self-improvement loop

the kit that gets better at itself
ai-error-learner
Catalogs recurring errors as AI pain points. The "we keep making this mistake" detector.
skill-builder
Transforms identified pain points into working skills. Pattern → implementation.
skill-improver
Monitors skill success rates and rewrites the ones under 80%. QA for your own AI.
How the loop works

Claude's mistakes become Claude's new skills.

01
ai-error-learner
Detects the same error three times. Files it as a pain point.
→
02
skill-builder
Reads the pattern. Writes a new skill that prevents the class of error.
→
03
skill-improver
Monitors the new skill. If it's under 80% success, rewrites it.
Powered by four Python hooks that track every tool call, error, and skill invocation.
Hooks

Automation at the tool-call layer.

Hooks fire around every tool call — before, after, and on stop. Block dangerous commands, auto-format after edits, run tests before Claude finishes. You write them in Python or bash; ClaudeKit ships templates for the common ones.

Hook types

PreToolUse
Before tool runs. Block dangerous commands, validate inputs, gate sensitive paths.
PostToolUse
After tool completes. Auto-format edited files, run linters, update metrics.
Stop
Before Claude finishes. Run tests, enforce quality gates, block incomplete work.

Included templates

format.sh Auto-format code after edits
validate-bash.sh Block dangerous bash commands
block-secrets.sh Prevent writes to sensitive files
skill_suggester.py Suggest skills from prompt keywords
security_gate.py Block edits to sensitive files
error_detector.py Track recurring errors
skill_monitor.py Track skill success rates
Install

Three ways to get started.

Pick whichever fits your workflow. All three land you with the same kit — memory, commands, hooks, and skills ready to use.

Option 1 — Recommended

Fastest path if you're starting a new repo.

  1. Visit the claudekit repo and click Use this template.
  2. Clone your new repository locally.
  3. Open with Claude Code — the context system is wired up and ready.
Option 2 — One-liner for existing projects

Run from the root of an existing repo. The installer prompts before overwriting any existing .claude/ or CLAUDE.md.

$ curl -sL https://raw.githubusercontent.com/Nnnsightnnn/claudekit/main/install.sh | bash
Option 3 — Manual

For when you want to see what you're getting before it lands in your repo.

# Clone and copy $ git clone https://github.com/Nnnsightnnn/claudekit.git /tmp/claudekit $ cp -r /tmp/claudekit/.claude /path/to/your/project/ $ cp /tmp/claudekit/CLAUDE.md /path/to/your/project/ $ rm -rf /tmp/claudekit
Already have ClaudeKit installed? Update with /update-template in Claude Code, or run ./update.sh from your repo root. Your memory, pain-points, and CLAUDE.md are preserved. Only commands and skill definitions are refreshed.