convez.ai
Skill
darthrater78

claude-vibe-skills

darthrater78/claude-vibe-skills

Ver no GitHub
InicialPluginScore · de 0 a 100

Two linked Claude Code skills: a gate-based vibe coding workflow enforcer and a security + Python best-practice guardrail

Instalação

$npx convez-ai darthrater78/claude-vibe-skills

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
22
Forks
2
Acompanhando
0
Saúde
28%
Último commit
há 1 semana
Versão
v2026.11.03
Issues abertas
2
Contribuidores
1

README

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

claude-vibe-skills

Three Claude Code skills: a cost-discipline mode, a gate-based release workflow enforcer, and a security guardrail — working together to keep vibe coding sessions cheap, safe, and properly shipped.

Skills

SkillFileDescription
dev-modedev-mode.skillOne-command launcher — loads all three skills simultaneously
cost-savercost-saver.skillCost-discipline mode — bounds output, gates model to Sonnet 4.6, trims MCP overhead
vibe-coding-workflowvibe-coding-workflow.skillGate-based release workflow enforcer
vibe-secure-vibe-codingvibe-secure-vibe-coding.skillSecurity & Python best-practice guardrail

Install all four and invoke /dev-mode (or say "dev mode") to activate the full suite at once. vibe-coding-workflow and vibe-secure-vibe-coding are also linked: the workflow automatically invokes the security skill at Gate 3 after every successful build.


Download

Download both .skill files from the latest release and install them (see below).

FileDirect download
dev-mode.skillDownload
cost-saver.skillDownload
vibe-coding-workflow.skillDownload
vibe-secure-vibe-coding.skillDownload

Using with Claude Desktop (claude.ai chat)

Skills are natively supported in the Claude Desktop app (Windows and macOS). The Desktop app runs an embedded skills plugin that automatically loads .skill files and makes them available in every conversation.

Install

  1. Download vibe-coding-workflow.skill and vibe-secure-vibe-coding.skill from the latest release
  2. Open Claude Desktop
  3. Click Customize in the sidebar
  4. Click Import skill and select each .skill file
  5. Confirm both skills appear as enabled

After importing, fully close and restart Claude Desktop before the skills will be available. Claude runs in the system tray after the window is closed — make sure to quit it from the taskbar/system tray icon too, not just close the window. Once restarted, skills activate automatically based on their trigger phrases — you don't invoke them manually. Just talk to Claude normally and they load when relevant (e.g. mention "vibe coding" or "audit my project").

Both skills must be installed together. vibe-coding-workflow calls vibe-secure-vibe-coding by name at Gate 3 — installing only one will cause Gate 3 to fail.

How it works

Skills are ZIP archives containing a SKILL.md with a name, trigger description, and instructions. Claude Desktop reads the trigger description and loads the skill into context whenever the conversation matches. The skill then guides Claude's behavior for that session.


Using with Claude Code (local CLI)

Claude Code is the terminal CLI (claude) that runs in your local environment. It shares the same skills registry as Claude Desktop — skills you install in the Desktop app are automatically available in Claude Code sessions too.

Install

Install both .skill files via Claude Desktop as described above. They will immediately be available in all Claude Code sessions on the same machine.

How it works in Claude Code

In Claude Code, skills appear in the system context at session start. When your conversation matches a skill's trigger, Claude loads it and follows the instructions. You'll see the skill name referenced in tool call outputs (e.g. Launching skill: vibe-coding-workflow).

Skills work the same way in Claude Code as in Desktop chat — the only difference is your working directory is your local codebase, so the workflow and security gates operate on real files and git state.

Using a skill directly from a file

In Claude Code you can also reference a .skill file directly without installing it:

claude "@path/to/vibe-coding-workflow.skill build and release this project"

This loads the skill for that session only, without permanently installing it.


Ad hoc use — drop into any chat or code session

You don't need to install a skill to use it. In both Claude Desktop chat and Claude Code, you can drop a .skill file directly into the session and it loads immediately for that conversation only.

Claude Desktop — drag into chat

Drag one or both .skill files into the chat input box, then type your prompt alongside them:

[drag vibe-coding-workflow.skill]  [drag vibe-secure-vibe-coding.skill]
build and release this project

Claude reads the skill files as context and follows their instructions for the rest of the session. Nothing is permanently installed — the skills only apply to that conversation.

Claude Code — @ reference in your prompt

Use the @ file syntax to reference a .skill file inline:

claude "@./vibe-coding-workflow.skill @./vibe-secure-vibe-coding.skill build and release this project"

Or interactively, type @ followed by the file path in the Claude Code prompt:

> @/path/to/vibe-coding-workflow.skill let's ship this

Again, session-only — nothing is installed.

When to use ad hoc vs installed

Ad hoc (drop in)Installed
Persists across sessionsNoYes
Auto-triggers from phrasesNo — must be in the promptYes
Requires install stepNoYes
Good forTrying a skill, one-off use, sharing with a teammateDaily use, automatic enforcement

Using without Claude Desktop (browser claude.ai)

Skills are not natively supported in the browser version of claude.ai. As a workaround, you can paste the skill content directly into your conversation:

  1. Download the .skill file

  2. Open it as a ZIP archive (rename to .zip or use any zip tool)

  3. Copy the contents of SKILL.md

  4. Paste it at the start of your Claude conversation with a note like:

    "Please follow these instructions for this session: [paste SKILL.md content]"

This gives Claude the same instructions for that session, though it won't persist across conversations and won't auto-trigger — you'll need to paste it each time.


dev-mode

One-command launcher that activates all three skills simultaneously. Use this instead of loading each skill individually at the start of a session.

Triggers

  • Invoke /dev-mode explicitly
  • Say "dev mode", "load dev mode", or "start dev mode"

What it loads

  1. cost-saver — model gating, output bounding, MCP overhead reporting, handoff prompts
  2. vibe-coding-workflow — 5-gate release process (version → build → security → release → push)
  3. vibe-secure-vibe-coding — inline security review + project audit mode

After loading, gives a single combined status summary and asks "What are we building?"

Install note

dev-mode also ships as a slash command for Claude Code CLI users. Copy dev-mode.md to ~/.claude/commands/dev-mode.md and invoke it with /dev-mode in any session — no skill install required.


cost-saver

Cost-discipline mode. Keeps sessions cheap without losing quality by targeting the five biggest cost drivers: session length, model choice, effort level, output length, and context bloat.

Triggers

Claude loads this skill automatically when you:

  • Mention cost, token usage, or expensive sessions
  • Say "keep it cheap", "save tokens", or similar
  • Ask to disable or trim MCP servers, connectors, or unused tools
  • Ask for a handoff summary to start a fresh session
  • Ask which model or effort level a task needs
  • Invoke /cost-saver

What it does

Bounds its own output. Leads with the answer, skips preamble and recaps, shows diffs instead of whole files, reads only the files a task requires.

Model gating — Sonnet 4.6 ceiling. At session start, checks the active model. If it's above Sonnet (Opus, Fable), immediately flags it:

"This session is running on [model], which exceeds the Sonnet 4.6 cost ceiling. Run /model sonnet to switch down, or tell me you want to stay on [model] for this task."

Won't proceed with substantial work until you respond. Can also write "model": "sonnet" to ~/.claude/settings.json so every new session defaults to Sonnet.

Phase-transition handoffs. When work shifts phase (explore → implement → test → done), offers a fresh-session prompt:

"Good point to start a fresh session — this one's carrying N messages of history. Want a handoff summary?"

If yes, produces a structured summary under 30 lines (goal / state / key files / decisions / next step) designed to replace the long history with a cheap restart.

MCP overhead reporting. Distinguishes between:

  • Active tools (full schemas loaded): expensive — often thousands of tokens per server per request. Named and flagged for removal.
  • Deferred tools (name-only, loaded on demand): low overhead — reported as a count only, no action needed.

For active servers not used in the current work, tells you exactly how to remove them (claude mcp remove, disabledMcpjsonServers in settings, Desktop UI toggle, or claude --strict-mcp-config for a zero-MCP session).

Effort fit. Recommends /effort high or xhigh only for architecture, root-cause debugging, or security review. Flags medium as sufficient for routine edits.

End-of-task token estimate. When a task wraps up, closes with 3–5 lines: per-request overhead × turns, files read vs. used, model multiplier, and the single highest-impact change for next time.

Token impact (rough estimate):
✅ Saved ~40k — read only 2 relevant files instead of exploring the package
✅ Saved ~25k/turn — session restarted at the implement phase
⚠️ ~30k/turn overhead — 5 connected MCP servers, none used this session
Biggest win next time: disable unused connectors (would cut ~60% of this session's tokens)

vibe-coding-workflow

Enforces a mandatory 5-gate release process for every vibe coding session. Gates must be cleared in order — none can be silently skipped.

🔢 VERSION  →  🔨 BUILD  →  🔒 SECURITY  →  📦 RELEASE  →  🚀 PUSH

Triggers

Claude loads this skill automatically when you:

  • Mention "vibe coding" or start a new coding session
  • Ask to build, compile, package, or bundle
  • Ask to push, deploy, release, or publish
  • Say "ship it", "done", or "ready to merge"
  • Try to skip a step ("just push it", "skip the version bump", "we can do security later")

The five gates

Gate 1 — Version 🔢

No build starts until a version is decided and written to the version file (package.json, pyproject.toml, Cargo.toml, VERSION, etc.). Must follow semver (MAJOR.MINOR.PATCH).

If no version is set:

🚫 VERSION GATE BLOCKED
No version has been set for this build. What version should this be?
  Patch bump (bug fixes):      x.x.N+1
  Minor bump (new features):   x.N+1.0
  Major bump (breaking):       N+1.0.0
Current version: [shown]

Gate 2 — Build 🔨

Runs the project's build command after version is confirmed. Build failure stops everything — no gate advances until the build is clean.

Gate 3 — Security 🔒

After every successful build, vibe-secure-vibe-coding is invoked automatically for a full project audit. This gate is mandatory — it cannot be deferred.

  • Critical and High findings are fixed immediately before proceeding
  • Medium and Low findings are surfaced for the user to decide on

Gate passes only when Critical = 0 and High = 0.

Gate 4 — Release 📦

Every security-cleared build must be formally published before the session ends. Claude tags the commit and creates a GitHub release (or equivalent) with notes and artifacts. "We'll release later" is blocked.

Gate 5 — Push 🚀

Claude will never push without an explicit typed confirmation. "yeah" or "ok" is not enough — the user must type something unambiguous like "push" or "confirm push".

Handling impatience

What you sayWhat happens
"just push it"Gates checked, push gate surfaced, confirmation required
"skip the version bump"Hard stop — version is required
"we can do security later"Security gate runs now, in this session
"just ship it"All open gates surfaced in order
"done" / "we're done"Release and push gates checked

vibe-secure-vibe-coding

Security and Python best-practice guardrail. Operates in two modes: inline review (as you write code) and project audit (full codebase scan).

Triggers

Claude loads this skill when you:

  • Write or review any code
  • Add packages or dependencies
  • Build APIs, handle user input, store credentials
  • Work with databases, files, or network operations
  • Say "just make it work", "hardcode it for now", "don't worry about security", "I'll fix it later"
  • Say "audit my project", "scan this codebase", "security review", "check my code"

What it checks

Security patterns

CategoryExamples caught
Secrets & credentialsHardcoded API keys, tokens, passwords; .env in git
Dangerous executioneval(), exec(), shell=True with user input, os.system()
Input handlingMissing validation, type/length/format checks
DatabaseSQL string concatenation (use parameterized queries)
NetworkDisabled TLS verification, SSRF-prone redirects
File systemPath traversal, overly permissive file modes
Serializationpickle/marshal on untrusted data
JavaScriptPrototype pollution, innerHTML XSS, open redirects

Python best practices

PatternRule
Type hintsEvery function signature should be annotated
Context managerswith for all file/DB/network/lock resources
pathlibUse Path instead of string concatenation for paths
secrets moduleUse for tokens, nonces, session IDs — not random
assert misuseNever use assert for validation — stripped by -O
Bare exceptAlways catch specific exception types
LoggingNever log passwords, tokens, or PII
HashingSHA-256+ for integrity; bcrypt/argon2 for passwords — never MD5/SHA1
Mutable defaultsdef f(x=None) not def f(x=[])
Pinned depsExact versions in requirements.txt; run pip audit in CI
Static analysisbandit in pre-commit or CI

Audit mode

Say "audit my project" or "scan this codebase" to trigger a full scan. Output format:

Security & Code Quality Audit — <project>
════════════════════════════════════════════════════════════

🚨 CRITICAL  — must fix before any production use
⚠️  HIGH      — fix before shipping
📝 MEDIUM    — fix soon, low exploitation risk
💡 LOW       — best-practice improvements

────────────────────────────────────────────────────────────
🚨 CRITICAL (N)
  <file>:<line>  <issue>
                 Fix: <one-line fix>
...

────────────────────────────────────────────────────────────
Summary
  Files scanned:   N
  Total findings:  N  (Critical: N  High: N  Medium: N  Low: N)

Severity guide

SeverityExamples
🚨 CriticalHardcoded secrets, SQL injection, shell=True with user input, disabled TLS, pickle on untrusted data
⚠️ HighPath traversal, missing auth, debug=True in prod, MD5/SHA1 passwords, bare random for tokens
📝 MediumBare except, no type hints, mutable defaults, assert for validation, unpinned deps
💡 LowMissing encoding= on open(), string path concatenation (non-traversal), missing bandit in CI

"Just make it work" handling

When asked to skip security, the skill doesn't lecture — it offers the safe version first (usually just as fast), and if the user insists, marks the shortcut visibly:

# SECURITY RISK: hardcoded credentials — move to env var before any real use
# TODO: os.environ['DB_PASSWORD']
DB_PASSWORD = "hunter2"

How they work together

vibe-coding-workflow
       │
       │  Gate 3 — Security
       └──invokes──► vibe-secure-vibe-coding
                            │
                            └── full project audit
                                fix Critical + High
                                surface Medium + Low
                                return findings report
                            │
       ◄──────────────────────
       │
       │  Gate 3 passes (0 Critical, 0 High)
       └──► Gate 4 — Release

vibe-secure-vibe-coding also runs independently whenever you're writing code — it doesn't require the workflow to be active.


Version

v1.3.0

Ver o original no GitHub

Bom para

Relacionados