convez.ai
SkillDesenvolvimento WebAvançado
ngxccc

Agent Skills Kit

ngxccc/agent-skills-kit

Harness spec-driven para agentes IA com memória contextual persistente

Ver no GitHub
BomPluginScore · de 0 a 100

Framework de desenvolvimento orientado por specs para agentes IA que gera automaticamente PRDs, gerencia backlogs e mantém memória contextual auto-melhorável. Permite que agentes de código trabalhem autonomamente por horas em tarefas grandes sem perder estado, gerando artefatos compartilháveis entre devs, PMs e stakeholders.

Instalação

$npx convez-ai ngxccc/agent-skills-kit

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
85%
Último commit
há 3 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Desenvolvimento autônomo de features com agentes IA mantendo specs atualizadas
  • Gerenciamento de backlogs e PRDs gerados automaticamente por agentes
  • Preservação de contexto técnico entre sessões longas de coding com IA

Veredito

Ferramenta sofisticada para equipes que usam agentes IA em desenvolvimento, resolvendo o problema real de perda de contexto em longas sessões. Requer expertise em automação e workflows com IA para aproveitar plenamente — não é plug-and-play.

README

A tradução para o português deste README ainda está na fila.

Built by world-class engineers, for agent-skillsrs at flowser.ai — AI Agents with computers for GTM

agent-skills-kit

This meta harness turns any AI coding agent into a spec-driven engineering team that researches, plans, ships production-grade code, and self-improves its memory to survive context-rotting even 6 months later.

🔬 Spec-driven development for AI agents 📋 Auto-generates PRDs, manages backlogs, routes context automatically 🧠 Self-improving knowledge base that compounds as you ship ⚡ Runs autonomously for hours on large tasks without losing state 🤝 Plans and specs are shareable — devs, PMs, and stakeholders review the same artifacts


🚀 Install (30 seconds)

Run this inside your project folder. Open a terminal and cd into the project you want the harness installed into before running the command — it installs into the current directory.

Prefer to drive it from your agent? Open Claude Code or Codex with that project folder as the working directory, then paste the full setup prompt below.

curl -fsSL https://raw.githubusercontent.com/ngxccc/agent-skills-kit/main/install.sh | bash

Then open Claude Code and say:

Run ag-setup

That's it. The setup skill detects your stack, asks about your project (a real conversation, not a checklist), scaffolds the process directory, deep-scans your codebase, and populates context files with actual content — not placeholders.

your-project/
├── .claude/
│   ├── agents/              # 🤖 12 specialized agent definitions
│   │   ├── ag-research-agent.md
│   │   ├── ag-execute-agent.md
│   │   └── ...
│   ├── skills/              # ⚡ 31 auto-discovered skills
│   │   ├── ag-generate-plan/
│   │   ├── ag-security/
│   │   ├── ag-scout/
│   │   └── ...
│   └── hooks/               # 🪝 7 lifecycle hooks
│       ├── privacy-block.cjs
│       ├── scout-block.cjs
│       └── ...
├── .codex/
│   └── agents/              # 🔄 Mirrored agents for Codex
├── CLAUDE.md                # 📋 Orchestrator + routing rules
├── AGENTS.md                # 📖 Agent registry
└── process/                 # 🧠 Created by ag-setup (not install)
    └── ...
  • Fresh project? Installs the full harness, then ag-setup studies your codebase
  • Existing .claude/ config? Backs up to .agent-skills-backup/, installs fresh, restores your settings.json
  • Existing process/ directory? Never touched by install — ag-setup handles migration intelligently
  • Existing CLAUDE.md? Backed up as CLAUDE.md.pre-agent-skills, harness version installed

First, open Claude Code or Codex with your project folder as the working directory (launch it from inside the project, or cd there first). The harness installs into the current directory, so this must be your project — then paste the prompt below.

First, install the agent-skills-kit agent harness by running this command:

curl -fsSL https://raw.githubusercontent.com/ngxccc/agent-skills-kit/main/install.sh | bash

After the install completes, run ag-setup to configure everything for this project.

Follow the full interactive flow:

1. DETECT — Read package.json, detect my stack (framework, package manager, monorepo
   structure, test framework, database, auth). Also check if I have any existing .claude/,
   process/, or context files from a previous setup.

2. SHOW ME WHAT YOU FOUND — Present a summary of the detection results and wait for me
   to confirm before continuing. If this is an existing project with process/ folders or
   context files, tell me what you found and what looks good vs what could be improved.

3. ASK ME ABOUT THE PROJECT — Before scaffolding or scanning, have a real conversation
   with me about this project. Don't just ask a fixed list of questions and move on — ask
   follow-ups based on my answers, probe deeper on anything vague, and keep going until
   you genuinely understand the project. Start with the basics (what is this? who uses it?),
   then dig into architecture, features, conventions, pain points, and anything else that
   matters. Summarize your understanding back to me and confirm it's correct before moving on.

4. SCAFFOLD — Create the process/ directory structure. If I already have process/ folders,
   show me what you plan to change and wait for my approval before reorganizing anything.
   Never silently move or delete my existing files.

5. STUDY — Deep-scan the codebase and populate process/context/all-context.md with REAL
   content based on what you find AND what I told you. Include: repo structure, tech stack
   with versions, key patterns and conventions, import aliases, env vars, API routes,
   database schema, test setup. Do not leave placeholder text.

6. VALIDATE — Run all the validation checks to make sure everything is wired correctly.

Important rules:
- If I have existing context files or a well-written CLAUDE.md, read them first and
  preserve what is good. Merge intelligently — do not replace good content with generic scans.
- Show me a summary of what you plan to create or change at each major step and wait
  for my OK before proceeding.
- Do not create empty placeholder files. Only create files that have real content.
- Ask before reorganizing. If my existing setup works, tell me what you would improve
  and let me decide.

🔥 The Problem

You ask Claude to "add webhook support." It immediately starts writing code. No questions about your architecture. No check on existing patterns. No plan. You get 400 lines that don't fit your codebase, and you spend an hour fixing it.

But that's just the surface. The deeper problems hit harder:

Your agent has intelligence but no process, no memory, and no way to collaborate with your team.

Whether you're a developer, a PM, or a CEO who just started vibe coding — this problem hits everyone the same way. The fix is the same too: give your agent a real development process.


🛠️ The Fix

This harness installs a complete development system into your project — not just a CLAUDE.md file, but 12 specialized agents, 31 skills, and a phase-locked workflow that forces your agent to understand before it builds.

%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '16px', 'lineColor': '#8888AA'}} }%%
flowchart TD
    R["🔍 RESEARCH\nRead codebase, gather facts"]
    I["💡 INNOVATE\nExplore 2-3 approaches"]
    P["📋 PLAN\nWrite detailed spec"]
    E["⚡ EXECUTE\nImplement the plan"]
    T["✅ tester → reviewer → git-manager"]
    U["🧠 UPDATE PROCESS\nCapture learnings"]

    R -->|"you say 'go'"| I
    I -->|"you say 'go'"| P
    P -->|"ENTER EXECUTE MODE"| E
    E --> T
    E -->|"recommended"| U

    style R fill:#1565C0,stroke:#0D47A1,color:#FFFFFF
    style I fill:#E65100,stroke:#BF360C,color:#FFFFFF
    style P fill:#2E7D32,stroke:#1B5E20,color:#FFFFFF
    style E fill:#C62828,stroke:#B71C1C,color:#FFFFFF
    style T fill:#6A1B9A,stroke:#4A148C,color:#FFFFFF
    style U fill:#00695C,stroke:#004D40,color:#FFFFFF

Every transition requires your explicit approval. Nothing auto-advances. You stay in control.


The Vibe Coding Revolution

Vibe coding changed who can build software. Spec-driven development changes what they can ship.

Most tools help you start a project. This harness helps you finish it — with plans your team can review, context that never goes stale, and safety systems that catch mistakes before they ship.


Who Is This For?

Whether you just discovered vibe coding or you're a staff engineer shipping production systems — this harness adapts to your workflow.


At a Glance


💎 Why Teams Use This

Most harnesses give you a CLAUDE.md and instructions. This gives you an autonomous development system that compounds intelligence over time.

📋 Spec-Driven Development — Not Vibes-Driven

Every feature gets a written plan with blast radius analysis before a single line of code is written.

💡 Auto-generates PRDs, manages backlogs, organizes feature groups. Works for both developers and product managers — your agent plans like a senior engineer, not an intern.

What's in every plan:

SectionPurpose
📍 TouchpointsEvery file that will be created or modified, listed upfront
📜 Public contractsWhat API surfaces or interfaces change
💥 Blast radiusWhat could break, what tests to run, what to watch
✅ Verification evidenceHow to prove the implementation is correct
🔄 Resume handoffEnough context for any agent to pick up mid-plan

🔄 Autonomous Multi-Phase Execution — Hours of Hands-Free Work

For large tasks, the agent runs an iterative phased loop:

🔍 research → ⚡ execute → ✅ validate → 📄 report → 🔄 repeat

💡 It self-heals when stuck, self-reflects to improve approach, and writes durable progress reports to disk. Context compaction can't kill it — all state lives in files, not memory.

Walk away and come back to completed work.

🧬 Auto-Architecture Research — Learn From Any Codebase

The agent doesn't just read your code — it studies other repositories to learn how they solved similar problems (ag-xia).

💡 It researches, compares approaches, and adapts the best patterns into your codebase. Architecture decisions are informed by real-world implementations, not hallucinated best practices.

🧭 Persistent Smart Context Routing — Always the Right Context

Context isn't one giant file. It's organized into auto-routed knowledge domains:

process/context/
├── all-context.md              # 🧭 Root router — reads your task, loads what's relevant
├── tests/
│   └── all-tests.md            # 🧪 Test runners, commands, debugging
├── container/
│   └── all-container.md        # 🐳 Docker, deployment, infra
├── uxui/
│   └── all-uxui.md             # 🎨 Components, design tokens, patterns
└── {your-domain}/
    └── all-{domain}.md         # 📚 Any domain with 3+ durable docs

💡 When the agent works on billing, it loads billing context — not your entire codebase docs. Context auto-updates every time you complete a feature, so it never goes stale.

🧠 Self-Improving Knowledge Base — Gets Smarter as You Ship

Every completed feature feeds learnings back into the context system.

💡 Research findings, architectural decisions, debugging insights, and coding patterns are captured and indexed automatically. Your 100th feature benefits from everything learned in the first 99. The knowledge compounds — it doesn't reset.


How This Compares

Featureagent-skills-kitSuperpowersGSDgstack
Spec-driven lifecycleFull RIPER-5 (research → plan → execute → verify)Mandatory workflowsContext-rot fixPartial
Phase-locked safetyTool restrictions per mode (read-only research, no-write innovate)Skill-based constraintsPhase separationNone
Multi-tool support7 tools via AGENTS.md + nativeClaude Code plugin14 runtimes1 tool
Auto-improving contextDomain-routed context groups, updates after every featurePlugin memoryDisk-persisted stateManual
Team collaborationShared specs, plans, and review artifactsSoloSoloSolo
Skills system32 auto-discovered, keyword-matched at every prompt86 composable skillsMeta-prompting23 role tools
Multi-phase programsUmbrella plans + phase-by-phase execution loop with regression checksSingle taskSingle taskSingle task
Quality pipeline6-step chain (code-review → test → simplify → security → audit → commit)Per-skill qualityNo auto-chainNo auto-chain
Installation30-second curl install + auto-setupPlugin marketplacenpx one-linergit clone
Context routingDomain-based routing table with grouped context packsFlat skill contextFlat contextSingle file

On runtime breadth: GSD supports 14 runtimes. We support 7 deeply — with full agent harnesses, skill discovery, and lifecycle hooks on every platform. Breadth vs. depth: your choice.


⚡ What Makes This Different

Most agent harnesses give you a big CLAUDE.md and some instructions. Here's what this one actually does:


🧭 How It Works

Your request
  → Step 0: Skill Discovery (match keywords → surface relevant skills)
  → Intent Detection (feature / bug / question / refactor / UI)
  → Route to the right agent
  → Phase-locked execution with explicit transitions

The orchestrator never does the work itself — it routes, monitors, and manages transitions.

📊 The Workflow

PhaseWhat happensYou say
🔍 RESEARCHRead-only fact gathering — codebase + web(automatic on feature requests)
💡 INNOVATEExplore 2-3 approaches with trade-offsgo
📋 PLANWrite a detailed spec you can reviewgo
⚡ EXECUTEImplement exactly what was plannedENTER EXECUTE MODE
🧠 UPDATE PROCESSCapture learnings, update context, archive plan(recommended after non-trivial work)

💡 Shortcuts: ENTER FAST MODE - [task] compresses RESEARCH+INNOVATE+PLAN into one pass — still pauses before EXECUTE. Trivial fixes (single file, <15 lines, no schema/auth changes) skip straight to execute.

💻 Typical Session

# 🆕 Feature request
You: "add webhook support to the API"
→ Skill discovery surfaces: ag-scenario, ag-security
→ research-agent gathers context (read-only, can't touch code)
→ You say "go" → innovate-agent explores approaches
→ You say "go" → plan-agent writes spec with blast radius
→ You review the plan, say "ENTER EXECUTE MODE"
→ execute-agent implements → self-review → tester → code-reviewer → git-manager
→ Closeout packet: what changed, what's verified, recommended next step
# 🐛 Bug fix
You: "login redirect is broken"
→ Routes to ag-debugger → evidence gathering → competing hypotheses
→ Root cause identified with proof chain
→ execute-agent implements the fix → quality pipeline
# ⏩ Fast mode
You: "ENTER FAST MODE - add rate limiting middleware"
→ Compressed research+innovate+plan in one pass
→ Mandatory safety pause → you review → "ENTER EXECUTE MODE"
# 🏗️ Large program
You: "build a full testing platform"
→ Creates umbrella plan + phase plans in a feature folder
→ Each phase: re-research → approve → execute → validate → durable report
→ Progress survives context compaction — durable reports on disk
# 🔄 Autonomous optimization
You: "improve test coverage to 80% using ag-autoresearch"
→ Agent iterates: make change → commit → measure → keep/revert
→ Stuck detection after 5 consecutive discards → strategy shift
→ Full audit trail in TSV

🛡️ Built-in Safety Systems

These aren't just guidelines — they're structural enforcement built into every agent.


🔍 Pre-Implementation Intelligence

Before a single line of code is written, the system can catch issues through specialized analysis:


🤖 Autonomous Agent Capabilities


✅ Quality Pipeline — Built Into Execution

The execute-agent doesn't just write code and call it done. It auto-chains through a quality pipeline:

%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '16px', 'lineColor': '#8888AA'}} }%%
flowchart TD
    E["⚡ Execute-Agent\nImplements the plan"]
    SR["🔎 Self-Review\nLine-by-line check\nagainst plan"]
    T["🧪 Tester\nDiff-aware — only\nruns affected tests"]
    CR["🔍 Code Reviewer\nEdge case scout\n+ adversarial review"]
    CS["✨ Code Simplifier\nClarity refactoring"]
    GM["📦 Git Manager\nLogical commit splitting\nfrom touched_files"]

    E --> SR
    SR --> T
    T --> CR
    CR --> CS
    CS --> GM

    style E fill:#C62828,stroke:#B71C1C,color:#FFFFFF
    style SR fill:#AD1457,stroke:#880E4F,color:#FFFFFF
    style T fill:#6A1B9A,stroke:#4A148C,color:#FFFFFF
    style CR fill:#283593,stroke:#1A237E,color:#FFFFFF
    style CS fill:#00695C,stroke:#004D40,color:#FFFFFF
    style GM fill:#37474F,stroke:#263238,color:#FFFFFF
StepWhat it does
🔎 Self-reviewChecks every checklist item against plan for deviations, documents them
🧪 TesterMaps changed files to test files, auto-escalates to full suite when >70% mapped
🔍 Code reviewerDispatches edge case scout BEFORE review, checks N+1 queries, auth paths, data leaks
✨ SimplifierClarity refactoring after review passes — no behavior changes
📦 Git managerReceives touched_files list, splits into logical conventional commits, refuses unknown files

📋 The Plan Lifecycle — Spec-Driven, Not Vibes-Driven

Every non-trivial feature follows a plan lifecycle — a written spec that is created, reviewed, executed against, and archived as project history.

%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '16px', 'lineColor': '#8888AA'}} }%%
flowchart TD
    A["🆕 Feature Request"]
    B["📝 Plan Created\nin active/"]
    C{"👀 User Reviews\nthe Plan"}
    D["⚡ Execute Against Plan"]
    E["📦 Plan Archived\nto completed/"]
    F["🧠 Learnings Written\nto all-context.md"]
    G["🔄 Next Feature\nStarts Smarter"]

    A --> B
    B --> C
    C -->|"✅ Approved"| D
    C -->|"✏️ Needs Changes"| B
    D --> E
    E --> F
    F --> G
    G -.->|"context compounds"| A

    style A fill:#1565C0,stroke:#0D47A1,color:#FFFFFF
    style B fill:#E65100,stroke:#BF360C,color:#FFFFFF
    style C fill:#F57F17,stroke:#F9A825,color:#000000
    style D fill:#C62828,stroke:#B71C1C,color:#FFFFFF
    style E fill:#6A1B9A,stroke:#4A148C,color:#FFFFFF
    style F fill:#00695C,stroke:#004D40,color:#FFFFFF
    style G fill:#1565C0,stroke:#0D47A1,color:#FFFFFF

💡 Six months from now, when someone asks "why did we build auth this way?", the answer is in completed/. Not lost in a Slack thread.

Where plans live on disk:

process/
├── general-plans/
│   ├── active/                  # 📋 Plans currently being worked on
│   │   └── webhooks_PLAN_28-05-26.md
│   ├── completed/               # ✅ Archived plans (searchable history)
│   ├── backlog/                 # 📌 Deferred work
│   ├── reports/                 # 📄 Cross-cutting reports
│   └── references/              # 📚 Research outputs
└── features/
    └── billing/                 # 🏷️ Feature-scoped (5+ artifacts)
        ├── active/
        ├── completed/
        ├── backlog/
        ├── reports/
        └── references/

🏗️ Phase Programs — Large Projects That Don't Fall Apart

Normal features use one plan. Large multi-phase projects use a phase program — an umbrella plan plus individual phase plans, each with its own validation gate.

%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '16px', 'lineColor': '#8888AA'}} }%%
flowchart TD
    UP["🎯 Umbrella Plan\nOverall program goal"]
    P1["📋 Phase 1 Plan"]
    P2["📋 Phase 2 Plan"]
    P3["📋 Phase 3 Plan"]

    R1["🔍 Re-Research"]
    E1["⚡ Execute"]
    V1["✅ Validate"]
    RP1["📄 Durable Report"]

    R2["🔍 Re-Research"]
    E2["⚡ Execute"]
    V2["✅ Validate"]
    RP2["📄 Durable Report"]

    UP --> P1
    UP --> P2
    UP --> P3

    P1 --> R1
    R1 -->|"approval"| E1
    E1 --> V1
    V1 --> RP1
    RP1 -.->|"learnings feed\nnext phase"| R2

    P2 --> R2
    R2 -->|"approval"| E2
    E2 --> V2
    V2 --> RP2

    style UP fill:#1565C0,stroke:#0D47A1,color:#FFFFFF
    style P1 fill:#E65100,stroke:#BF360C,color:#FFFFFF
    style P2 fill:#E65100,stroke:#BF360C,color:#FFFFFF
    style P3 fill:#E65100,stroke:#BF360C,color:#FFFFFF
    style R1 fill:#1565C0,stroke:#0D47A1,color:#FFFFFF
    style E1 fill:#C62828,stroke:#B71C1C,color:#FFFFFF
    style V1 fill:#2E7D32,stroke:#1B5E20,color:#FFFFFF
    style RP1 fill:#6A1B9A,stroke:#4A148C,color:#FFFFFF
    style R2 fill:#1565C0,stroke:#0D47A1,color:#FFFFFF
    style E2 fill:#C62828,stroke:#B71C1C,color:#FFFFFF
    style V2 fill:#2E7D32,stroke:#1B5E20,color:#FFFFFF
    style RP2 fill:#6A1B9A,stroke:#4A148C,color:#FFFFFF

Key features:

FeatureWhy it matters
🔄Re-research at every phaseChecks for code drift, reads latest reports, updates assumptions
✅Validation gatesPhase isn't VERIFIED until evidence proves it. Honest status: PLANNED → CODE DONE → TESTING → VERIFIED or BLOCKED
📄Durable reportsEvery phase writes results to disk. Progress survives context compaction
🧠Learnings feed forwardPhase 1 discoveries update Phase 2's plan before execution
🏗️Foundation vs expansionExplicitly splits "prove the architecture" from "implement everything"
🚧Honest blocker handlingBlocked phases stay BLOCKED with evidence. No forcing green status

🧠 Context Groups — Organized Knowledge, Not One Giant File

Project knowledge is organized into context groups — durable knowledge domains, each with an all-{group}.md router that tells agents what to read and when.

process/context/
├── all-context.md              # 🧭 Root router — architecture, stack, patterns, conventions
├── tests/
│   └── all-tests.md            # 🧪 Test runners, commands, debugging procedures
├── container/
│   └── all-container.md        # 🐳 Docker, deployment, infra procedures
├── uxui/
│   └── all-uxui.md             # 🎨 Components, design tokens, patterns
├── infra/
│   └── all-infra.md            # 🖥️ Worker nodes, provisioning, DNS
├── skills/
│   └── all-skills.md           # ⚡ Skill runtime, app architecture
├── workflows/
│   └── all-workflows.md        # 🔄 Workflow runtime, deployment
└── {your-domain}/
    └── all-{domain}.md         # 📚 Any knowledge domain with 3+ durable docs
How it works
🧭 Router patternAgents read only what's relevant to their task, not everything
📏 Auto-promotionTopics with 3+ docs or 800+ lines get their own context group
🔄 Living docsUpdated by update-process-agent after every non-trivial feature
🧪 Auditableag-audit-context verifies routing and consistency

📁 Feature Folders — Self-Organizing Project Memory

When a topic accumulates 5+ artifacts, it gets its own feature folder — a complete lifecycle container.

process/features/{feature}/
├── active/       # 📋 Plans currently being worked on
├── completed/    # ✅ Archived plans (searchable decision history)
├── backlog/      # 📌 Deferred work (agents check before duplicating)
├── reports/      # 📄 Execution reports, post-mortems, validation results
└── references/   # 📚 Research outputs that inform future decisions
What happens
🆕New work starts in active/ → reports accumulate → plan archives to completed/
📌Deferred work goes to backlog/ — agents check it before creating duplicate plans
📦Feature promotion happens automatically when general artifacts hit 5+
🔍Every feature has complete, self-contained history — plans, decisions, reports, research

🤖 What's Inside

12 Agents

Core workflow agents — one per RIPER-5 phase:

AgentRole
🔍 ag-research-agentCodebase + web research, read-only. Contradiction tracking built in
💡 ag-innovate-agentBrainstorm 2-3 approaches. Must produce decision summary before PLAN
📋 ag-plan-agentWrite spec with anti-rationalization guards. "I already know how" is not a plan
⚡ ag-execute-agentImplement per plan. 50% check-in, deviation protocol, self-review
⏩ ag-fast-mode-agentCompressed RESEARCH→INNOVATE→PLAN with mandatory safety pause
🧠 ag-update-process-agent7-phase mandatory checklist including stale artifact scanning

Specialist agents — called during EXECUTE or standalone:

AgentRole
🐛 ag-debuggerEvidence-before-hypothesis. Competing hypotheses, elimination chains
🧪 ag-testerDiff-aware. Only runs affected tests. Auto-escalates on config changes
🔎 ag-code-reviewerEdge case scout BEFORE review. N+1 detection, auth path validation
✨ ag-code-simplifierClarity refactoring without behavior change
🎨 ag-ui-ux-designerDesign-aware frontend. Can spawn research subagent mid-execution
📦 ag-git-managerLogical commit splitting from touched_files. Refuses unknown files

51 Skills (auto-discovered)

🔧 Contract skills — ag-generate-plan · ag-generate-context · ag-audit-context · ag-audit-plans · ag-audit-ag · ag-setup · ag-update · ag-publish

🧠 Planning — ag-predict (5-persona debate) · ag-scenario (12-dimension edge cases) · ag-sequential-thinking · ag-problem-solving

🐛 Debug & security — ag-debug · ag-security (STRIDE + OWASP + auto-fix) · ag-autoresearch (autonomous optimization) · ag-upstash-ratelimit

📚 Research — ag-docs-seeker · ag-scout · ag-docs · ag-repomix · ag-xia (repo comparison)

🎨 Frontend — ag-frontend-design · ag-chrome-devtools · ag-agent-browser · ag-web-testing · ag-nextjs · ag-nextjs-bundle-optimizer · ag-nextjs-ppr-patterns · ag-react-hook-form · ag-react-hook-form-audit · ag-tailwind · ag-tailwind-refactor · ag-tailwind-responsive-ui · ag-tailwind-ui-refactor · ag-ui-design

⚙️ Utilities — ag-context-engineering · ag-mcp-management · ag-preview · ag-team (parallel agents) · ag-tech-graph · ag-watzup (session handoff) · ag-merge-worktree · ag-zod · ag-demo-irc · ag-code-simplifier · ag-implementation-design-patterns · ag-implementation-functional-patterns · ag-typescript · ag-typescript-advanced-patterns · ag-typescript-refactor · ag-skill-standard

💡 Some skills (like ag-xia) were inspired by ClaudeKit by @mrgoonie. We focused on fewer, deeper skills instead of 80+.

🪝 7 Hooks

HookWhat it does
🔐 Privacy guardrailsBlocks .env, credentials, SSH keys. Requires explicit approval
🚫 Scout blockerPrevents agent from wandering into node_modules/, dist/. Gitignore-syntax .ckignore
🧠 Session initDetects stack, injects env vars, recovers approval gates after compaction
💉 Subagent contextInjects ~200 token compact context block into every subagent
✨ Edit qualityAfter 5+ edits, nudges to run code-simplifier (non-blocking, throttled)
📛 Descriptive namingLanguage-aware file naming conventions on every Write
📊 Usage trackingSession metrics and token awareness

Where everything lives:

your-project/
├── .claude/
│   ├── agents/              # 🤖 12 agent definitions (.md)
│   ├── skills/              # ⚡ 31 skill modules (each a directory with SKILL.md)
│   └── hooks/               # 🪝 7 lifecycle hooks (.cjs)
├── .codex/
│   └── agents/              # 🔄 Mirrored for Codex compatibility
├── .agents/
│   └── skills -> ../.claude/skills   # 🔗 Symlink for Codex discovery
├── CLAUDE.md                # 📋 Orchestrator config + routing rules
├── AGENTS.md                # 📖 Agent + skill registry
└── process/
    ├── context/             # 🧠 Auto-routed knowledge domains
    ├── general-plans/       # 📋 Cross-cutting plans + reports
    ├── features/            # 🏷️ Feature-scoped lifecycle folders
    └── development-protocols/  # 📜 Shared workflow rules

🔄 Updating

Pull the latest harness improvements:

Run ag-update

💡 Shows a dry-run diff, waits for confirmation. Your process/ directory and project-specific content are never touched.


Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

Quick links:

🙏 Credits

This project is cloned and customized from vibecode-pro-max-kit by @withkynam.

This project benefited from ClaudeKit by @mrgoonie — particularly skills like ck:xia which inspired some of ours.

The difference: agent-skills-kit focuses on the spec-driven development framework and self-improving context organization, without bloating you with 80+ skills. Fewer tools, more structure.


⭐ Star History


📄 License

MIT

Ver o original no GitHub

Bom para

Relacionados