Skip to content
/

centminmod/my-claude-code-setup

centminmod-cc-setup · centminmod/my-claude-code-setup · ★ 2.4k · last commit 2026-05-24

Provides a memory-resilient Claude Code starter with dual git-shared + machine-local memory that survives CLAUDE.md resets, plus multi-provider delegation to Codex and Z.AI.

Best whenMemory bank files should be gitignored from project commits but version-controlled privately, separating project context from AI memory artifacts.
Skip ifSpeculating about code you haven't read, Committing AI changes without user confirmation (vibe coding)
vs seeds
superpowers(skills + hook), but centminmod-cc-setup uniquely implements a dual-memory architecture (git-shared CLAUDE-*.md + machin…
Primitive shape 27 total
Commands 14 Skills 6 Subagents 6 Hooks 1
00

Summary

centminmod/my-claude-code-setup — Summary

centminmod/my-claude-code-setup is a personal Claude Code starter template built and maintained by George Liu (centminmod), shipping three CLAUDE.md template variants, a dual-memory architecture, a small command/skill/agent library, and a macOS notification hook. It is explicitly a "template repo" — users fork or copy files into their own projects rather than installing a package. The dual-memory system stores git-shared CLAUDE-*.md memory bank files alongside machine-local memory/ auto memory files, providing resilience against CLAUDE.md resets. The repo includes six subagents (code-searcher, codex-cli, memory-bank-synchronizer, ux-design-expert, zai-cli, get-current-datetime), six skills, and ~24 slash commands organized into thematic subdirectories. A unified_notifier.py Stop hook fires desktop notifications via terminal-notifier on macOS. Supplementary CLAUDE-cloudflare.md, CLAUDE-convex.md, and CLAUDE-history-template.md docs provide platform-specific context injection for Cloudflare, Convex, and ClerkOS projects.

differs_from_seeds: Closest to agent-os (markdown-scaffold + per-domain context files) and superpowers (skills-only plus hooks). Unlike agent-os, it ships substantive slash commands (24+) and named subagents with Chain-of-Draft mode support. Unlike superpowers it has no iron-law behavioral framework — the rules file is concise and scoped. Its unique differentiator is the dual-memory architecture (git-shared CLAUDE-*.md + machine-local auto-memory with explicit sync workflow), which no seed implements, combined with a Z.AI integration allowing GLM-4.7 model access alongside Claude.

01

Overview

centminmod/my-claude-code-setup — Overview

Origin

Personal Claude Code configuration published by George Liu (centminmod) at https://github.com/centminmod/my-claude-code-setup. Active since at least early 2025 with a significant v2026 update on May 4, 2026 that introduced three modernized CLAUDE.md templates and a migration guide. The repo is maintained by one primary contributor with 2,359 GitHub stars.

Philosophy

The project's philosophy is captured in its CLAUDE.md:

"This is a template repo — no application code. Users fork/copy files to their own projects."

"Be sure to read the official Claude Code docs first... Sign up for a paid Claude AI account."

The core insight is memory resilience: CLAUDE.md can be reset or wiped by Claude Code during compact operations; the dual-memory architecture ensures project knowledge survives via a parallel machine-local memory/ directory that mirrors key CLAUDE-*.md content. The README documents the full memory recovery workflow:

"If CLAUDE.md is ever reset or wiped, auto memory retains project knowledge — check /memory to recover context."

Rules are kept to a minimum in .claude/rules/core-rules.md:

"Never speculate about code you have not read. Read files and ripgrep for usages before making claims."

"Do what has been asked; nothing more, nothing less."

Target User

Claude Code users on macOS (the notification hook uses terminal-notifier), particularly those building web apps with Cloudflare, Convex, or using Z.AI as an alternative provider for higher token quotas and GLM-4.7 model access.

Alternative READMEs

The author had Claude Code generate multiple README variants (Progressive Disclosure, Task-Based, Technical Reference) to serve different reader types — itself a demonstration of the tool's practical use.

02

Architecture

centminmod/my-claude-code-setup — Architecture

Distribution

Standalone GitHub repository. No install script, no npm package. Users manually copy or fork files to their project directories.

Install Method

# Manual copy approach
git clone https://github.com/centminmod/my-claude-code-setup
cp -r .claude/ /path/to/your-project/
cp CLAUDE.md /path/to/your-project/
# Optional: install terminal-notifier for macOS hooks
brew install terminal-notifier

Directory Tree (repo root)

centminmod/my-claude-code-setup/
├── .claude/
│   ├── agents/                  (6 named subagents)
│   ├── commands/                (8 subdirs of slash commands)
│   │   ├── anthropic/           (3 commands)
│   │   ├── architecture/        (2 commands)
│   │   ├── ccusage/             (1 command)
│   │   ├── cleanup/             (1 command)
│   │   ├── documentation/       (1 command)
│   │   ├── promptengineering/   (2 commands)
│   │   ├── refactor/            (1 command)
│   │   └── security/            (3 commands)
│   ├── hooks/
│   │   ├── unified_notifier.py       (macOS notification hook)
│   │   └── unified_notifier.py.md    (hook documentation)
│   ├── mcp/
│   │   └── chrome-devtools.json      (Chrome DevTools MCP config)
│   ├── rules/
│   │   └── core-rules.md             (behavioral rules)
│   ├── skills/
│   │   ├── ai-image-creator/
│   │   ├── audit-session-metrics/    (haiku-pinned, cheapness optimization)
│   │   ├── claude-docs-consultant/
│   │   ├── consult-codex/
│   │   ├── consult-zai/
│   │   └── session-metrics/
│   ├── settings.json                 (Stop hook, model: sonnet)
│   └── settings.local.json
├── CLAUDE.md                         (Template 3 format — progressive disclosure)
├── CLAUDE-template-1.md              (~101 lines, compact)
├── CLAUDE-template-2.md              (~153 lines, memory bank)
├── CLAUDE-template-3.md              (~105 lines, progressive disclosure)
├── CLAUDE-cloudflare.md              (Cloudflare platform docs)
├── CLAUDE-cloudflare-mini.md         (condensed Cloudflare docs)
├── CLAUDE-convex.md                  (Convex database docs)
├── CLAUDE-history-template.md        (session history template)
├── CLAUDE-migrate-to-new-template.md (migration prompt)
├── AGENTS.md
├── GEMINI.md                         (Gemini CLI companion)
└── mcp-servers.md                    (MCP config reference)

Required Runtime

  • Claude Code (Claude.ai paid subscription)
  • macOS (for terminal-notifier hook; hook can be removed for Linux/Windows)
  • Optional: terminal-notifier (brew install terminal-notifier)
  • Optional: ripgrep (brew install ripgrep), fd, jq

Target AI Tools

Primary: Claude Code. Also references Z.AI (GLM-4.7 via Codex-compatible API), Codex CLI, and provides a GEMINI.md for Gemini CLI compatibility.

Config Files

  • .claude/settings.json — Stop hook, model selection (sonnet), cleanupPeriodDays: 365
  • .claude/settings.local.json — machine-local overrides
  • .mcp.json (in user projects) — Chrome DevTools MCP reference provided

Context Layers Architecture

Layer Location Loads Shared Resilient
Project context CLAUDE.md Always Git No
Core rules .claude/rules/core-rules.md Always Git Yes
Auto memory memory/MEMORY.md Always (200 lines) No Yes
Auto memory topics memory/*.md On demand No Yes
Path-scoped rules .claude/rules/*.md Matching files Git Yes
Skills .claude/skills/ On demand Git Yes
03

Components

centminmod/my-claude-code-setup — Components

Slash Commands (24 total across 8 subdirectories)

anthropic/ (3)

  • apply-thinking-to.md — applies Anthropic extended thinking patterns to a prompt file; uses progressive reasoning structure, multi-phase enhancement methodology
  • convert-to-todowrite-tasklist-prompt.md — converts task descriptions into TodoWrite-style structured task lists
  • update-memory-bank.md — syncs CLAUDE-*.md memory bank files and auto memory topic files after significant work

architecture/ (2)

  • explain-architecture-pattern.md — identifies and explains architectural patterns (MVC, Clean Architecture, DDD, etc.) with ASCII diagrams
  • check-best-practices.md — audits code against architecture best practices

ccusage/ (1)

  • ccusage-daily.md — displays daily Claude Code token usage statistics

cleanup/ (1)

  • cleanup-context.md — optimizes CLAUDE.md context, removes stale entries

documentation/ (2)

  • create-readme-section.md — generates README sections
  • create-release-note.md — drafts release notes from git history

promptengineering/ (2)

  • batch-operations-prompt.md — structures prompts for batch file operations
  • convert-to-test-driven-prompt.md — converts implementation prompts to TDD format

refactor/ (1)

  • refactor-code.md — refactors code preserving behavior; includes test verification gates

security/ (3)

  • secure-prompts.md — audits prompts for injection vulnerabilities
  • security-audit.md — full security audit of codebase
  • test-examples/ — security test prompt examples

Agents (6)

  • code-searcher.md — comprehensive codebase search and analysis; supports Chain of Draft (CoD) ultra-concise mode; model: sonnet, color: purple
  • codex-cli.md — delegates tasks to Codex CLI for GPT-family model execution
  • get-current-datetime.md — returns current date/time (utility agent)
  • memory-bank-synchronizer.md — syncs memory bank files across sessions
  • ux-design-expert.md — UI/UX design guidance and code review
  • zai-cli.md — delegates tasks to Z.AI (GLM-4.7 access via Z.AI API)

Skills (6)

  • ai-image-creator/ — generates images via AI APIs
  • audit-session-metrics/ — audits session-metrics JSON exports; pinned to Haiku for cost optimization; supports quick/detailed modes; operates on session/project/instance JSON scopes
  • claude-docs-consultant/ — fetches and interprets Claude Code documentation
  • consult-codex/ — routes sub-tasks to Codex CLI; enables multi-model delegation
  • consult-zai/ — routes sub-tasks to Z.AI GLM-4.7 models
  • session-metrics/ — collects and exports session token usage data

Hooks (1)

  • unified_notifier.py — Python script triggered on Stop event; sends macOS desktop notification via terminal-notifier: "Finished working in $(basename $PWD)". Requires macOS + terminal-notifier.

MCP Servers (1, reference config only)

  • chrome-devtools.json — Chrome DevTools MCP server configuration; not bundled, reference-only

Templates / Memory Docs

  • CLAUDE-template-1.md — compact (~101 lines)
  • CLAUDE-template-2.md — memory bank headline + dual memory (~153 lines)
  • CLAUDE-template-3.md — progressive disclosure (~105 lines)
  • CLAUDE-cloudflare.md — Cloudflare + ClerkOS platform reference
  • CLAUDE-cloudflare-mini.md — condensed Cloudflare reference
  • CLAUDE-convex.md — Convex + Next.js + Cloudflare Pages reference
  • CLAUDE-history-template.md — session history pattern
  • CLAUDE-migrate-to-new-template.md — AI prompt for CLAUDE.md migration

Rules

  • .claude/rules/core-rules.md — behavioral rules: investigation accuracy, scope discipline, verification/safety, efficiency, memory resilience
05

Prompts

centminmod/my-claude-code-setup — Prompts

Verbatim Excerpt 1: audit-session-metrics SKILL.md (haiku-pinned skill)

---
name: audit-session-metrics
model: haiku
description: >
  Audit a session-metrics JSON export for token-usage waste and produce a
  plain-English findings report. Trigger when the user runs
  /audit-session-metrics, when session-metrics suggests an audit after an
  HTML export, or when the user asks to audit / review / find waste in a
  saved session-metrics JSON. Two modes: "quick" (ratios + cache health +
  top expensive turns/sessions) and "detailed" (adds CLAUDE.md / settings /
  re-read scan). Supports session, project, and instance JSON scopes.
  Args: $ARGUMENTS[0] = quick|detailed,
  $ARGUMENTS[1] = path to a session-metrics JSON export.
---

# Audit Session-Metrics

Reads a session-metrics JSON export and produces a prioritised, plain-English
audit of token-usage waste. Pinned to Haiku in frontmatter so it runs ~10×
cheaper than Sonnet for a turn that is mostly summarisation work.

Prompting technique: Model-pinning via YAML frontmatter (model: haiku). Explicit cost optimization at the prompt level — chooses a cheaper model because the task is summarization, not generation. Uses $ARGUMENTS[0]/$ARGUMENTS[1] parameter routing to branch behavior without separate commands.


Verbatim Excerpt 2: apply-thinking-to.md command

You are an expert prompt engineering specialist with deep expertise in applying
Anthropic's extended thinking patterns to enhance prompt effectiveness. Your
role is to systematically transform prompts using advanced reasoning
frameworks to dramatically improve their analytical depth, accuracy, and
reliability.

**ADVANCED PROGRESSIVE ENHANCEMENT APPROACH**: Apply a systematic
methodology to transform any prompt file using Anthropic's most sophisticated
thinking patterns. Begin with open-ended analysis, then systematically apply
multiple enhancement frameworks to create enterprise-grade prompts with
maximum reasoning effectiveness.

**TARGET PROMPT FILE**: $ARGUMENTS

## SYSTEMATIC PROMPT ENHANCEMENT METHODOLOGY

### Phase 1: Current State Analysis & Thinking Pattern Identification

<thinking>
I need to thoroughly analyze the current prompt to understand its purpose,
structure, and existing thinking patterns before applying enhancements.
What type of prompt is this? What thinking patterns would be most beneficial?
What are the specific enhancement opportunities?
</thinking>

### Phase 2: Sequential Enhancement Framework Application

Apply these enhancement frameworks:
- Framework 1: Progressive Reasoning Structure
  Before: "Analyze the code for security issues"
  After: "Before applying standard security frameworks, think creatively about
  what unique security characteristics this codebase might have..."

Prompting technique: Meta-prompt (a prompt about improving prompts). Uses <thinking> tags explicitly to demonstrate extended thinking patterns. Employs Before/After transformation examples as few-shot demonstrations of the enhancement technique itself.


Verbatim Excerpt 3: code-searcher.md agent (Chain of Draft patterns)

## Chain of Draft Few-Shot Examples

### Example 1: Finding Authentication Logic
**Standard approach (150+ tokens):**
"I'll search for authentication logic by first looking for auth-related
files, then examining login functions..."

**CoD approach (15 tokens):**
"Auth→glob:*auth*→grep:login|jwt→found:auth.service:45→implements:JWT+bcrypt"

### Key CoD Patterns:
- **Search chain**: Goal→Tool→Result→Location
- **Error trace**: Bug→Search→Line→Cause→Fix
- **Architecture**: Pattern→Structure→Components→Framework
- **Abbreviations**: impl(implements), fn(function), cls(class), dep(dependency)

Prompting technique: Chain of Draft (CoD) — few-shot demonstration of ultra-compressed reasoning chains. Token counting per approach (15 vs 150+ tokens) as explicit optimization metric. Mode detection via keyword scanning.

09

Uniqueness

centminmod/my-claude-code-setup — Uniqueness

differs_from_seeds

Closest to agent-os (markdown scaffold, minimal automation) and superpowers (skills + single hook). Unlike agent-os, centminmod ships substantial slash commands (24+) and named subagents, not just CLAUDE.md templates. Unlike superpowers, there is no iron-law behavioral framework — rules are concise and non-dogmatic. The unique architectural differentiator is the dual-memory system: git-shared CLAUDE-*.md files plus machine-local memory/ auto-memory with explicit sync workflow. No seed in the reference set implements this pattern. The second differentiator is multi-provider delegation: consult-codex and consult-zai skills route sub-tasks to OpenAI Codex and Z.AI (GLM-4.7), giving Claude Code users access to alternative LLMs within a single session — a pattern not present in any of the 11 seeds. Compared to ccmemory (Neo4j-backed memory), centminmod uses plain markdown with no infrastructure dependency.

Positioning

"My Claude Code starter settings" — a practitioner's published personal config, not an enterprise framework. Documentation quality and multiple README variants (4 formats for different reader types) suggest high user-empathy. The author documents their own AI journey (Substack, Threads, BlueSky), positioning this as a teaching resource.

Observable Failure Modes

  1. macOS-only hook: unified_notifier.py requires terminal-notifier (macOS only). On Linux/Windows the Stop hook fails silently (or errors out). Users must delete or replace the hook.
  2. Manual memory sync discipline: The dual-memory system requires the user to remember to run /update-memory-bank after each significant session. There is no automatic trigger. If users forget, memory diverges.
  3. No install script: All file copying is manual. Configuration drift is likely when the upstream repo updates.
  4. Z.AI dependency: The consult-zai and zai-cli components require a Z.AI API key and account. Users without Z.AI access get broken agents.
  5. Dev container URL: References https://claude-devcontainers.centminmod.com/ which may become stale.

Cross-References

  • References Z.AI integration guide: https://z.ai/subscribe?ic=WWB8IFLROM
  • References OpenClaw security: https://github.com/centminmod/explain-openclaw
  • Author writes supplementary AI content at https://ai.georgeliu.com
04

Workflow

centminmod/my-claude-code-setup — Workflow

Setup Phases

Phase Action Artifact
1. Copy files Fork/copy .claude/ + CLAUDE.md to project root Project-local .claude/
2. Select template Choose CLAUDE-template-1/2/3 or migrate existing CLAUDE.md Customized CLAUDE.md
3. Run /init Claude Code analyzes codebase, populates memory bank CLAUDE-*.md memory files
4. Install hook deps brew install terminal-notifier (macOS) Working Stop notifications
5. Ongoing: update memory After significant work, run /update-memory-bank Updated CLAUDE-.md + memory/.md

Memory Sync Workflow

After significant work (refactors, new features, bug fixes):

  1. Run /update-memory-bank → Claude updates CLAUDE-activeContext.md, CLAUDE-patterns.md, CLAUDE-decisions.md, CLAUDE-troubleshooting.md
  2. Claude syncs key content to memory/MEMORY.md (index), memory/patterns.md, memory/architecture.md, memory/build.md
  3. If CLAUDE.md is ever reset: run /memory → recovers context from auto memory topic files

Memory Bank Files

File Purpose Loaded When
CLAUDE-activeContext.md Current session state and goals Session start
CLAUDE-patterns.md Code patterns and conventions Before implementing
CLAUDE-decisions.md Architecture Decision Records Before design choices
CLAUDE-troubleshooting.md Known fixes and debugging patterns When debugging
CLAUDE-config-variables.md Config values and env vars When touching config

Approval Gates

No formal approval gates defined. The CLAUDE.md template requires:

  • "Ask before destructive actions: deleting files/branches, force pushes, hard resets, --no-verify"
  • Showing diffs before committing

Artifacts Per Phase

Phase Produces
Session start Auto memory injection from memory/MEMORY.md
Task completion Updated CLAUDE-.md + memory/.md
Session end macOS notification (Stop hook)
Security audit /security-audit.md report
Release Release notes via /create-release-note
06

Memory Context

centminmod/my-claude-code-setup — Memory & Context

Dual-Memory Architecture

The repo's primary innovation is its two-layer memory system:

Layer 1: Git-Shared Memory Bank (CLAUDE-*.md files)

Files in the repo root, version-controlled, read on demand:

File Content Read When
CLAUDE-activeContext.md Current goals, session state Session start
CLAUDE-patterns.md Code patterns and conventions (numbered list) Before implementing
CLAUDE-decisions.md Architecture Decision Records (ADR-NNN format) Before design choices
CLAUDE-troubleshooting.md Known issue/fix pairs When debugging
CLAUDE-config-variables.md Config values, env vars When touching config

These files are gitignored from project commit history (per CLAUDE.md recommendation: "Exclude CLAUDE.md and CLAUDE-*.md from commits") to avoid polluting shared repos with AI memory files. Users maintain private backups via rsync.

Layer 2: Machine-Local Auto Memory (memory/*.md)

Not git-tracked, survives CLAUDE.md resets:

File Mirrors Lines
memory/MEMORY.md Index of all memory bank files 200 (always loaded)
memory/patterns.md CLAUDE-patterns.md Variable
memory/architecture.md CLAUDE-decisions.md Variable
memory/build.md Build, Test & Verify section Variable

Sync Workflow

After significant work:

  1. User runs /update-memory-bank command
  2. Claude updates CLAUDE-*.md files
  3. Claude syncs key content to memory/*.md topic files
  4. On recovery: user runs /memory → Claude reads from memory/*.md

Memory Type

File-based (markdown). No database, no vector store. Plain markdown files in repo root and memory/ subdirectory.

Persistence Scope

  • CLAUDE-*.md: project-scoped (per project directory), git-shared, not committed
  • memory/*.md: project-scoped, machine-local, never git-committed

Compaction Handling

The Stop hook fires at session end. The dual-memory system is the compaction recovery mechanism: if Claude Code compacts and loses CLAUDE.md, the memory/MEMORY.md auto-loads at 200 lines and provides recovery pointers.

Cross-Session Handoff

Yes — explicit. CLAUDE-activeContext.md documents current session state; user updates it before ending sessions. CLAUDE-history-template.md provides a structured session history pattern for long-running projects.

State Files Written

  • memory/MEMORY.md
  • memory/patterns.md
  • memory/architecture.md
  • memory/build.md
  • CLAUDE-activeContext.md
  • CLAUDE-patterns.md
  • CLAUDE-decisions.md
  • CLAUDE-troubleshooting.md
  • exports/session-metrics/ (from session-metrics skill)
07

Orchestration

centminmod/my-claude-code-setup — Orchestration

Multi-Agent Support

Yes — via 6 named subagents. No explicit orchestration protocol defined.

Orchestration Pattern

sequential — The main Claude thread dispatches to agents (code-searcher, codex-cli, zai-cli) as needed. No parallel fan-out, no queue management. The code-searcher agent explicitly supports Chain of Draft mode for token-efficient analysis. The consult-codex and consult-zai skills enable multi-model delegation.

Multi-Model Usage

Yes — unique feature:

  • Default model: Claude Code (Sonnet, per settings.json)
  • consult-codex skill: delegates to OpenAI Codex / GPT-family via Codex CLI
  • consult-zai skill: delegates to Z.AI GLM-4.7 models via Z.AI API
  • audit-session-metrics skill: pinned to haiku model for cost optimization
  • codex-cli.md agent: explicit Codex CLI routing
  • zai-cli.md agent: explicit Z.AI routing

This makes centminmod-cc-setup notable for multi-provider routing within a Claude Code session.

Isolation Mechanism

None (in-place file editing). No git worktree automation, no container isolation.

Execution Mode

Interactive-loop. No background daemon, no scheduled tasks.

Crash Recovery

None explicit. Dual-memory system provides CLAUDE.md reset recovery but not crash state recovery.

Context Compaction

Yes — via dual-memory system. Pre-compact and post-compact behavior not hooked; resilience is via CLAUDE-.md → memory/.md sync.

Cross-Session Handoff

Yes — CLAUDE-activeContext.md + memory/MEMORY.md.

Consensus Mechanism

None.

08

Ui Cli Surface

centminmod/my-claude-code-setup — UI / CLI Surface

Dedicated CLI Binary

No — no custom CLI binary. The repo ships files for users to copy into their projects. Claude Code itself is the runtime interface.

Local Web Dashboard

No — no web UI.

IDE Integration

VS Code + Claude Code extension recommended but not required. The repo references Claude Code VSC Extension in the README and provides an optional dev container setup.

Dev Container

Optional: Claude Code + Codex + Gemini CLI + OpenCode + Vercel CLI + Cloudflare Wrangler + Amazon AWS CLI in a single isolated VS Code dev container running Debian 12. Referenced at https://claude-devcontainers.centminmod.com/. Enables dangerously_skip_permissions for YOLO mode.

Notification / Observability

  • Stop hook (unified_notifier.py): macOS desktop notification via terminal-notifier when session ends
  • session-metrics skill: captures token usage statistics, exports to exports/session-metrics/ as JSON (session/project/instance scopes)
  • audit-session-metrics skill: reads session-metrics JSON exports, produces token-waste analysis reports
  • ccusage-daily command: displays daily usage statistics

Status Line

Not implemented — no custom status line.

Slash Command Access

24 commands across 8 thematic subdirectories accessible via /command-name or /<subdir>/command-name in Claude Code.

Git Automation

No automatic commits, PRs, or merges. CLAUDE.md template instructs: "Show the user what files are being committed, briefly explain the key changes, and ask for confirmation before proceeding."

Related frameworks

same archetype · same primary tool · same memory type

claude-mem (thedotmack) ★ 78k

Background worker service captures every tool call as an observation, AI-compresses sessions, and auto-injects relevant past…

pi (badlogic/earendil) ★ 55k

A minimal, hackable, multi-provider terminal coding agent that adapts to your workflows via npm-installable TypeScript Extensions…

Agent Skills (Addy Osmani) ★ 46k

Encodes senior-engineer software development lifecycle as 23 auto-routed skills and 7 slash commands for any AI coding agent.

wshobson/agents Plugin Marketplace ★ 36k

Single Markdown source for 83 domain-specialized plugins that auto-generates idiomatic artifacts for five AI coding harnesses.

TabbyML/Tabby ★ 34k

Self-hosted AI coding assistant server (alternative to GitHub Copilot) with admin dashboard, RAG-based completions, and multi-IDE…

Compound Engineering ★ 17k

Make each unit of engineering work compound into easier future work via brainstorm→plan→execute→review→learn cycles.