claude-impl-tools

> Implementation skill pack for Claude Code — Build software with AI agent teams

A plugin of 28 skills for Claude Code. Covers the full dev lifecycle: planning, orchestration, quality, security, maintenance, and skill ecosystem intelligence. Skills auto-install project-level hooks and agents on demand.

---

Quick Start

Option 1: Plugin Install (Recommended)

Launch Claude Code, then run:

/plugin marketplace add insightflo/claude-impl-tools
/plugin install claude-impl-tools@insightflo

Option 2: Quick Install (No plugin)

curl -fsSL https://raw.githubusercontent.com/insightflo/claude-impl-tools/main/scripts/quick-install.sh | bash

This clones the repo and symlinks skills. No hooks or agents installed — skills handle that on demand.

---

What You Get

| Component | Count | Installed when | |-----------|-------|----------------| | Skills | 28 | Plugin install | | Hooks | up to 17 | On demand (skill runs install.sh --local) | | Worker Agents | 4 | On demand (project-level) | | Templates | 11 | On demand (project-level) |

---

Skills (28)

Core Workflow

| Skill | What it does | |-------|--------------| | /workflow | Meta hub — analyzes project state, recommends the next skill (rules + experience hybrid) | | /agile | Layered sprints (Skeleton → Muscles → Skin) for 1-30 tasks | | /recover | Resume work after CLI crashes, network drops, or agent errors | | /checkpoint | Two-stage code review at task/PR completion |

Project Initialization

| Skill | What it does | |-------|--------------| | /si-planning | SI requirements engineering — Socratic elicitation, domain checklist, auto-generated deliverables (requirements analysis / functional spec / screen spec / traceability matrix), change impact analysis, customer dashboard | | /governance-setup | Set up Agent Teams governance or Mini-PRD | | /tasks-init | Generate TASKS.md from scratch | | /tasks-migrate | Consolidate legacy task files into TASKS.md |

Quality & Security

| Skill | What it does | |-------|--------------| | /quality-auditor | Comprehensive audit + verification discipline + quantitative metrics (v3.0) | | /security-review | OWASP Top 10, CVE, secrets detection | | /multi-ai-review | Claude + Gemini + Codex 3-AI consensus engine with domain presets |

Orchestration

| Skill | What it does | |-------|--------------| | /team-orchestrate | Unified engine — 3 modes: auto (Task dispatch), team (Agent Teams API), thin (ultra-minimal for 50-200 tasks) | | /cmux-orchestrate | cmux-based multi-AI team — Claude/Gemini/Codex in physical parallel | | /cmux-ai-run | cmux pane-split parallel task execution | | /cmux-ai-review | cmux pane-split parallel 3-stage review | | /multi-ai-run | Role-based AI model routing (Codex=code, Gemini=design, Claude=plan) | | /whitebox | Execution control plane — state, health, approvals | | /cmux | cmux terminal multiplexer control |

Maintenance & Analysis

| Skill | What it does | |-------|--------------| | /maintenance | ITIL 5-stage production maintenance orchestrator | | /impact | Blast radius and risk analysis before file modification | | /deps | Dependency graph + circular dependency detection | | /changelog | Change history query by domain/date/author | | /coverage | Test coverage analysis and gap identification | | /architecture | Project structure and tech stack visualization | | /compress | Context optimization with self-editing /prune (H2O + Context-1 pattern, v2.0) | | /statusline | TASKS.md progress in Claude Code status bar |

Skill Intelligence

| Skill | What it does | |-------|--------------| | /memento | Skill ecosystem engine — experience logging, smart routing, failure reflection, harness generation, cross-project knowledge via DuckDB |

---

Memento: Skill Ecosystem Intelligence

/memento learns which skills work best for which tasks and improves the ecosystem over time.

| Mode | Purpose | |------|---------| | log | Record skill execution outcomes (auto via PostToolUse hook) | | route <task> | Experience-weighted skill recommendation | | health | Unified skill dashboard across all projects | | reflect <skill> | Analyze failure patterns, suggest SKILL.md fixes | | harness <skill> | Generate validation scripts from failure patterns (AutoHarness-inspired) | | global search | Cross-project knowledge search via DuckDB | | global recall | Topic-specific cross-project learning retrieval |

Prerequisites for Memento

Memento's SKILL.md is included in the plugin. The execution infrastructure requires one-time global setup:

# 1. Install DuckDB
pip install duckdb

# 2. Initialize the global experience store
python3 -c "
import duckdb
db = duckdb.connect('$HOME/.claude/memento/experience.duckdb')
# Tables are auto-created on first use
db.close()
"

# 3. Add experience logging hook to ~/.claude/settings.json
# (See Harness Infrastructure below)

---

Orchestration Modes

/team-orchestrate auto-selects the right mode based on task count:

| Task count | Mode | How it works | |-----------|------|--------------| | 1-50 | --mode=auto | Direct Task dispatch, worktree-based phases | | 10-80 | --mode=team | Agent Teams API, 3-level hierarchy | | 50-200 | --mode=thin | Ultra-minimal context (76% token reduction) |

# Examples
/team-orchestrate                    # auto-selects mode
/team-orchestrate --mode=thin        # force thin mode
/auto-orchestrate                    # alias for --mode=auto

---

Agent Teams

Uses Claude Code native Agent Teams with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.

Your Claude session (= team lead)
├── architecture-lead
│     ├── backend-builder
│     └── reviewer
├── design-lead
│     ├── frontend-builder
│     └── designer
└── qa-lead

Enabling Agent Teams

bash plugin/project-team/install.sh --local --mode=team

Or let /team-orchestrate auto-install when needed.

Project-level hooks (installed on demand)

| Category | Hooks | |----------|-------| | Permission | permission-checker, domain-boundary-enforcer | | Safety | pre-edit-impact-check, risk-area-warning, security-scan | | Quality | standards-validator, design-validator, interface-validator, quality-gate | | Gates | policy-gate, contract-gate, risk-gate, docs-gate | | Sync | architecture-updater, changelog-recorder, cross-domain-notifier, task-sync |

Hook modes

| Mode | Hooks | Use case | |------|-------|----------| | lite | 4 | MVP, startups | | standard | 7 | Most projects | | full | 17 | Regulated industries | | team | 8 + Agent Teams hooks | Team orchestration |

---

Harness Infrastructure (Global)

The plugin works out of the box, but full harness engineering features require global setup in ~/.claude/:

Invariant Hooks (PreToolUse)

Enforce rules with code, not just instructions. Add to ~/.claude/settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Write|Edit",
        "hooks": [
          { "type": "command", "command": "bash ~/.claude/memento/invariants/cross-domain-guard.sh", "timeout": 3 },
          { "type": "command", "command": "bash ~/.claude/memento/invariants/sensitive-file-guard.sh", "timeout": 2 }
        ]
      },
      {
        "matcher": "Bash",
        "hooks": [
          { "type": "command", "command": "bash ~/.claude/memento/invariants/test-before-commit.sh", "timeout": 3 }
        ]
      }
    ],
    "PostToolUse": [
      {
        "matcher": "Skill",
        "hooks": [
          { "type": "command", "command": "bash ~/.claude/memento/log_skill.sh", "timeout": 5 }
        ]
      }
    ]
  }
}

Global files

| File | Purpose | |------|---------| | ~/.claude/memento/experience.duckdb | Unified experience store (DuckDB) | | ~/.claude/memento/query.py | CLI: search, health, recall, sync, sql | | ~/.claude/memento/log_skill.sh | PostToolUse hook — auto-log skill executions | | ~/.claude/memento/update_outcome.py | Update experience outcome with failure taxonomy | | ~/.claude/memento/invariants/*.sh | PreToolUse invariant enforcement scripts |

---

Recommended Workflow

Start
  │
  ├─ "What should I do?" ────────────── /workflow
  │
  ├─ Plan your project
  │   ├─ Large project? ──────── /governance-setup
  │   └─ Generate tasks ───────── /tasks-init
  │
  ├─ Implement (choose by scale)
  │   ├─ Small (≤30 tasks) ───── /agile auto
  │   ├─ Medium (30-50) ──────── /team-orchestrate --mode=auto
  │   ├─ Large (50-200) ──────── /team-orchestrate --mode=thin
  │   └─ With cmux ────────────── /cmux-orchestrate
  │
  ├─ Maintain
  │   ├─ Before editing ──────── /impact
  │   ├─ Bug fix ──────────────── /maintenance
  │   ├─ Check dependencies ──── /deps
  │   └─ Review changes ──────── /changelog
  │
  ├─ Quality
  │   ├─ Test coverage ───────── /coverage
  │   ├─ Security scan ───────── /security-review
  │   ├─ Pre-deploy audit ────── /quality-auditor
  │   └─ Context cleanup ─────── /compress (or /prune)
  │
  ├─ Skill ecosystem
  │   ├─ Skill performance ──── /memento health
  │   ├─ Cross-project recall ── /memento global recall
  │   └─ Improve a skill ─────── /memento reflect <skill>
  │
  └─ If interrupted ───────────── /recover

---

Repository Structure

claude-impl-tools/
├── .claude-plugin/
│   └── marketplace.json           # Marketplace manifest
├── plugin/
│   ├── .claude-plugin/
│   │   └── plugin.json            # Plugin manifest (version, metadata)
│   ├── skills/                    # 26 skills (auto-discovered)
│   │   ├── memento/               # Skill ecosystem intelligence
│   │   ├── team-orchestrate/      # Unified orchestration (3 modes)
│   │   ├── quality-auditor/       # Comprehensive audit (v3.0)
│   │   ├── context-optimize/      # H2O + self-editing (v2.0)
│   │   └── ...
│   └── project-team/              # On-demand project setup
│       ├── install.sh             # Local installer
│       ├── agents/                # Worker agents
│       ├── hooks/                 # Validation & governance hooks
│       └── templates/             # Protocols, ADR, contracts
└── scripts/
    └── quick-install.sh           # Alternative install (clone + symlink)

---

Requirements

| Component | Requirement | |-----------|-------------| | All skills | Claude Code CLI | | /memento global features | Python 3.10+, pip install duckdb | | /multi-ai-review, /cmux-ai- | gemini CLI, codex CLI (optional) | | /cmux- skills | cmux terminal multiplexer | | Project Team hooks | Node.js 18+ | | Agent Teams | CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 | | Invariant hooks | Global ~/.claude/memento/invariants/*.sh setup |

---

Version History

| Version | Changes | |---------|---------| | 4.7.0 | /si-planning skill (SI requirements engineering, domain profiles, customer dashboard), GSD benchmark integration | | 4.6.0 | GSD patterns: context monitoring hooks, /discuss skill, goal-backward verification, deviation rules, autonomous mode | | 4.5.0 | context-optimize v2.0 (self-editing /prune), hybrid recall | | 4.4.0 | Unified DuckDB experience store, memento global mode | | 4.3.1 | memento harness mode (AutoHarness-inspired validation) | | 4.3.0 | memento skill, team-orchestrate 3-mode unification, quality-auditor v3.0 | | 4.2.0 | cmux live-mode: paste-buffer + callback scripts | | 3.5.x | cmux skills, auto-revision quality improvements | | 3.4.0 | Agent Teams API migration |

---

License

---

English | 한국어

claude-impl-tools

Summary

Install to Claude Code