thanx-agent-starter

A starter kit for building your personal AI development ecosystem. Fork it, fill in your context, and start compounding.

What This Is

This repo gives you the self-improving infrastructure that makes AI-assisted development compound over time. It includes four core meta-skills and a structure for accumulating knowledge across sessions. The skills are not workflow-specific — they are the infrastructure that makes all your future skills get better.

The core skills and patterns originated from Darren Cheng's personal AI workflow, demonstrated at the Thanx AI Academy (Feb 2026). This repo generalizes that system into a forkable starting point.

Philosophy

• Fork and diverge. Clone this, make it yours. Your personal repo should reflect how YOU work.
• The improve loop is everything. Run /improve at the end of sessions. Each session leaves the system better than it found it.
• Start personal, promote when ready. Build skills for yourself first. When something proves universal, contribute it back to the core.
• Steal freely. Browse other people's forks for ideas. Copy what works. Adapt it. No permission needed.

The Three-Layer Model

Before you fork, understand where skills live. This mental model prevents the most common mistake — dumping everything into your personal repo and wondering why project-specific skills don't transfer.

┌─────────────────────────────────────────────────┐
│  Layer 3: Team-wide (thanx-claude-plugins)      │
│  Standards everyone follows. Shared reviews,    │
│  investigation workflows, ops patterns.         │
├─────────────────────────────────────────────────┤
│  Layer 2: Project-specific (.claude/commands/)  │
│  Skills tied to a specific codebase. Deploy     │
│  scripts, test runners, project conventions.    │
├─────────────────────────────────────────────────┤
│  Layer 1: Personal (this repo)                  │
│  YOUR workflow, preferences, automation.        │
│  /improve, /handoff, voice profile, knowledge.  │
└─────────────────────────────────────────────────┘

Start at layer 1. Build skills for yourself. When something proves useful beyond your personal workflow, promote it to layer 2 (project) or layer 3 (team). This repo is your layer 1.

Quick Start

# Fork and clone
gh repo fork thanx/thanx-agent-starter --clone
cd thanx-agent-starter

# Run setup (creates symlinks to ~/.claude/)
./setup.sh

# Fill in your context
$EDITOR global/CLAUDE.md      # Your identity, preferences, org context

Then start using Claude Code in any project. The skills are available everywhere via /disk-cleanup, /improve, /handoff, /knowledge, /write-skill, /sync-upstream, /browse-skills, and /register-repo.

What Is Included

Core Skills

Skill	Purpose
/disk-cleanup	Scan local disk for large storage consumers and identify cleanup opportunities. Read-only by default, never deletes without approval.
/improve	End-of-session learning loop. Extracts friction points, proposes skill improvements, captures knowledge. The compounding mechanism.
/handoff	Generates structured context prompts for transferring work between agent sessions or repos.
/knowledge	Manages a structured knowledge base for facts, patterns, and insights that persist across sessions.
/write-skill	Skill authoring guide with platform conventions, gotchas, and validation. Makes your first custom skill work correctly.
/sync-upstream	Checks the upstream scaffold for improvements and helps you incorporate them into your fork.
/browse-skills	Discover and adopt skills from the central skill ecosystem. Reads the repo-list, scans repos via GitHub API.
/register-repo	Register your skill repository in the central ecosystem repo-list via an automated PR.

Templates

File	Purpose
global/CLAUDE.md	Your global instructions, loaded into every conversation. Fill in your context.
global/settings.json	Default permission sandboxing. Allows reads, blocks destructive ops, prompts for everything else.
global/PERMISSIONS.md	Explains the allow/deny model and how to customize permissions.
context/voice-profile.md	Instructions for generating a writing voice profile from your emails and messages.
context/knowledge/index.md	Empty knowledge base registry. Grows as you capture learnings.

Repository Structure

thanx-agent-starter/
├── README.md                          # This file
├── CLAUDE.md                          # Instructions for working IN this repo
├── setup.sh                           # Creates symlinks to ~/.claude/
├── global/
│   ├── CLAUDE.md                      # Template -> symlinked to ~/.claude/CLAUDE.md
│   ├── settings.json                  # Permission defaults -> symlinked to ~/.claude/settings.json
│   └── PERMISSIONS.md                 # Explains the allow/deny model
├── agents/
│   └── skills/                        # Core skills -> symlinked to ~/.claude/skills/
│       ├── disk-cleanup/SKILL.md
│       ├── improve/SKILL.md
│       ├── handoff/SKILL.md
│       ├── knowledge/SKILL.md
│       ├── write-skill/SKILL.md
│       ├── sync-upstream/SKILL.md
│       ├── browse-skills/SKILL.md
│       └── register-repo/SKILL.md
├── community/
│   ├── repo-list.yaml                 # Central registry of known skill repos
│   └── README.md                      # Ecosystem docs
├── context/
│   ├── voice-profile.md              # Instructions for generating yours
│   └── knowledge/
│       └── index.md                   # Empty knowledge base registry
└── .gitignore

How It Works

The Symlink Model

The setup script creates three symlinks:

• ~/.claude/CLAUDE.md -> <your-repo>/global/CLAUDE.md
• ~/.claude/skills/ -> <your-repo>/agents/skills/
• ~/.claude/settings.json -> <your-repo>/global/settings.json

This means your global instructions, skills, and permission defaults are version-controlled in your fork, but Claude Code picks them up automatically in every project.

The Improve Loop

The core workflow:

1. Work in any project using Claude Code or Conductor
2. At the end of the session, run /improve
3. It analyzes the session, identifies friction, proposes skill improvements
4. Improvements get applied to your skills, knowledge gets captured
5. Next session starts with a better system

This is the compounding mechanism. Without it, skills are static markdown. With it, every session makes the system smarter.

Staying Current with Core

Run /sync-upstream from within this repo to check for improvements to core skills. It will:

1. Fetch the latest from the upstream scaffold
2. Show you what changed with a recommendation for each (adopt, skip, or review)
3. Help you merge selected changes without overwriting your personal content

You can also do it manually:

# Add upstream remote (one-time)
git remote add upstream git@github.com:thanx/thanx-agent-starter.git

# See what changed
git fetch upstream
git log HEAD..upstream/master --oneline

# Cherry-pick specific improvements, or ask Claude to help incorporate them

Generating Your Voice Profile

One of the most impactful early steps is generating a voice profile — a document that captures how you actually write so Claude can match your style.

See context/voice-profile.md for instructions. The short version: point Claude at your last 30 days of outgoing emails and Slack messages, and ask it to extract your writing patterns.

Skill Ecosystem

This repo hosts the central repo-list at community/repo-list.yaml — a discovery aid for the decentralized skill-sharing ecosystem across Thanx.

Three-Layer Discovery Model

1. Personal repos (kosmin-claude-skills, anutron/claude-skills, drn/dots) — each person's curated skills
2. Central repo-list (this repo: community/repo-list.yaml) — discovery aid listing known repos
3. Hub (thanx-cortex) — centralized, org-wide skills that benefit everyone

How to Participate

1. Fork this repo and add your own skills
2. Register your repo with /register-repo (creates a PR to add it to the repo-list)
3. Discover others' skills with /browse-skills (reads the repo-list, scans repos via GitHub API)
4. Adopt what you like into your personal repo
5. Centralize organically — when 3+ people share similar skills, promote to thanx-cortex

See community/README.md for full details.

Contributing Back

Found an improvement to a core skill that would benefit everyone? Open a PR against thanx/thanx-agent-starter. Good candidates for upstream contributions:

• Bug fixes in core skills
• New platform conventions discovered (Claude Code gotchas, dynamic context patterns)
• Structural improvements to the improve loop or knowledge capture
• Better template sections in the CLAUDE.md

NOT good candidates for upstream:

• Workflow-specific skills (expense reports, leadership prep, etc.)
• Org-specific context or cached IDs
• Personal preferences baked into core skills

License

Internal to Thanx. Fork freely within the organization.