vibecode-pro-max-kit
A spec-driven coding harness for vibecoders, bringing plan-first RIPER-5 workflows, self-healing loops, and persistent memory to any AI coding agent.
- Source repo
- withkynam/vibecode-pro-max-kit
- Stars
- ★ 1.2k
- Last updated
- 4mo ago
- License
- MIT
- Primary language
- JavaScript
- FA score
- 47/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformChatGPT · Codex · Claude Code
- Cost
- Free software; you pay for model usage
- Setup effort
- Low · running in minutes
- You'll need
- Typical use
- A CEO/founder wants to build a SaaS with auth, billing, and a landing page; the agent researches the stack, writes a reviewable architecture plan, implements with tests, and captures decisions for later audit.
- Not a fit if
- Windows users who need PowerShell or cmd.exe, not Git Bash or WSL
- Teams wanting a GUI or desktop app instead of terminal commands
- Beginners who want zero-config and no large file tree written into their project
- Source review
- 47/100 · Major gaps
What does this agent do, and when should you use it?
vibecode-pro-max-kit is a comprehensive developer tool that installs into any project to give AI coding agents a structured, plan-first development process. It comprises 15 specialized agents, 33 skills, 10 lifecycle hooks, and 22 protocols, enforcing a seven-phase workflow from research to execution. The kit introduces two self-healing loops—PVL (Plan-Validate-Fix) and EVL (Execute-Validate-Fix)—that automatically check and correct plans and tests, up to 10 cycles each. Autopilot mode enables hands-free runs that survive session resets by persisting state to disk, and the /goal command provides a copy-pasteable token for resuming autonomous runs. The kit is tool-agnostic, leveraging open standards (AGENTS.md and SKILL.md) to work across Claude Code, Codex, Cursor, Windsurf, and more. Installation is a single curl command that safely routes to interactive setup or update, requiring Node.js ≥ 22 and bash.
The kit installs a structured set of files into your project: .claude/agents/ (15 agent definitions), .claude/skills/ (33 skills), .claude/hooks/ (10 hooks), CLAUDE.md, AGENTS.md, and process/ directory with development protocols. It defines a strict RIPER-5 workflow: Research (read-only fact gathering) → Spec (requirements doc) → Innovate (2-3 approaches) → Plan (detailed checklist) → Validate (PVL loop) → Execute (implementation) → Update Process (capture learnings). At runtime, the coordinator agent performs skill discovery by scanning 33 skills and matching keywords, evaluates request ambiguity on a 4-signal scale, and routes to appropriate sub-agents like vc-research-agent, vc-plan-agent, and vc-execute-agent. The coordinator never writes code or runs tests itself—all execution happens in sub-agents, ensuring structural honesty. The PVL loop validates plans before coding (V1–V7 checkpoints), while the EVL loop independently re-runs tests after execution. Autopilot mode uses /goal blocks to enable hands-free, autonomous runs that can resume after session resets, with only three hard-stop conditions (irreversible actions, blocked phases, or paid probes).
- A CEO/founder wants to build a SaaS with auth, billing, and a landing page; the agent researches the stack, writes a reviewable architecture plan, implements with tests, and captures decisions for later audit.
- A product manager needs a dashboard showing MRR, churn, and growth metrics; the agent generates a PRD-style spec, gets approval, implements to spec, and archives the plan as project history.
- A designer wants pixel-perfect matching of a Figma screenshot; the design-aware agent analyzes the mockup, implements component-by-component with design tokens, and runs visual comparison checks.
- An engineer needs to refactor an auth module to support RBAC with zero downtime; the agent researches current code, writes a migration plan mapping affected files, and implements safely with rollback notes.
- A developer needs a quick one-line fix; fast mode skips heavy ceremony to keep the change simple.
- A team is building a large platform; umbrella plans and per-phase inner loops keep the project from falling apart.
How do you install or deploy this agent?
Prerequisites: Node.js ≥ 22, git, bash (macOS/Linux/WSL/Git Bash). Run this command inside your project folder:
curl -fsSL https://raw.githubusercontent.com/withkynam/vibecode-pro-max-kit/main/install.sh | bashThe installer detects if you're new or returning and prints the next step: for fresh projects, run claude and say "Run vc-setup"; for upgrades, say "Run vc-update". Codex users run /vc-setup or /vc-update.
How do you use this agent?
After installation, open your AI coding agent (e.g., Claude Code) with your project as the working directory and say "Run vc-setup". The setup flow detects your tech stack, asks about your project, creates the process/ directory, scans your codebase, and validates the setup. Then you can make requests like "add webhook support to the API". The agent performs skill discovery, clarifies intent if needed, routes to the research agent, and guides you through phases. Use "ENTER VALIDATE MODE" and "ENTER EXECUTE MODE" to progress. For hands-free runs, say autopilot full: [task] or paste a /goal block.
What are this agent's strengths and limitations?
- Plan-first RIPER-5 workflow prevents agents from jumping straight to code, reducing costly mistakes.
- Two independent quality loops (PVL and EVL) self-heal plans and tests, ensuring correctness before shipping.
- Autopilot mode runs hands-free for hours, survives session resets by persisting state to disk.
- Coordinator never inlines execution, making the process structurally honest—code only via sub-agents.
- Multi-tool compatibility via AGENTS.md and SKILL.md open standards, supporting 7+ coding assistants.
- Requires Node.js ≥ 22 and bash, with Windows users forced to use Git Bash or WSL; PowerShell not supported.
- Installation creates many files and directories, which may overwhelm beginners despite being non-destructive.
- Autopilot relies on reliable state files; corruption could cause resumption failures.
- Smart model policy uses expensive models (e.g., opus) for coding, potentially increasing costs.
- No detailed troubleshooting guide; users may rely on GitHub issues for help.
How does this agent compare with similar options?
Compared to Superpowers, GSD, and gstack, this kit offers a more comprehensive RIPER-5 lifecycle, two quality check loops (PVL + EVL), broader multi-tool support (7 tools), auto-improving knowledge, and team collaboration features. While GSD supports 14 runtimes, this kit focuses on depth rather than breadth.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| vibecode-pro-max-kit This agent | 47 · Major gaps | CLIFree + model costs | ★ 1.2k | 4mo ago | JavaScript | ChatGPT · Codex · Claude Code |
| RecallLoom Project Memory | Insufficient evidence | Agent plugin / skillFree | ★ 154 | 2mo ago | Python | Codex · Claude Code |
| Emulo | 78 · Good | Agent plugin / skillFree + model costs | ★ 297 | 16d ago | HTML | Codex · Claude Code |
| Citadel | 60 · Some gaps | Agent plugin / skillFree + model costs | ★ 924 | today | JavaScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows non-destructive install with backups, but least privilege not explicitly addressed; user confirmation exists in interactive setup, but autopilot may bypass; data flow transparency lacking; sensitive data handling mentioned but not detailed; dependency security not assessed, only pinned GitHub Actions; external effects limited but install runs external code; rollback exists via backups but untested; source attribution unclear, publisher unverified.
Self-consistency is good, README aligns with code structure, but minor inconsistencies (e.g., 33 skills vs 36 validators); dependency availability unverified, relies on Node.js and network; failure messages not detailed but install script provides guidance.
Broad audience (developers, PMs, CEOs), diverse scenarios; capability boundaries not explicit but skills layered; trigger precision high (/goal, /vc-setup); environment fit good (multi-platform, multi-tool).
Information architecture clear with TOC and docs; install notes detailed; naming stability moderate with reserved namespace; examples and FAQ present; known limitations mentioned (e.g., Windows symlink); license MIT; versioning and changelog exist but not detailed; maintenance responsibility unclear.
Output usability high, provides complete workflow; marginal value high, addresses context loss; cost-benefit not quantified but claims reduced model costs.
Claim traceability moderate, some claims supported by code; cross-source corroboration insufficient, relies on single repo; fact-inference separation unclear.
- Publisher unverified, source attribution unclear.
- Autopilot mode may bypass user confirmation, use with caution.
- Data flow transparency lacking, no details on data collection or transmission.
- Dependency security not assessed, install script executes external code.
Recent releases
Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.
-
v3.2.5 v3.2.5 — Windows install guidance
Release notes excerpt
Docs-only release. Documents the Windows install path (Git Bash / WSL). The installer already supported Windows shells (OSTYPE detection, symlink→copy fallback, LF-enforced shebangs); the gap was the README never mentioned Git Bash. Now the install prerequisites tell Windows users to run inside Git Bash or WSL — not PowerShell/cmd. No script changes.View on GitHub ↗ -
v3.2.4 v3.2.4: install.sh data-loss fix (defer content dirs to vc-update)
Release notes excerpt
Fixes a data-loss bug in the curl install.sh | bash upgrade path. The legacyDeletions pass previously rm -rf'd every listed directory unconditionally, including process-layout content dirs (process/general-plans/reports, process/general-plans/references, and _seeds equivalents). Since install.sh is deterministic (no agent), it cannot run the adaptive safe-migration vc-update Part D performs — so real reports/references content would be destroyed before vc-update ever ran. install.sh now defers any non-empty reports/references dir under process/ (preserved, not deleted) for vc-update …View on GitHub ↗ -
v3.2.3 v3.2.3: vc-update runs adaptive migration on version-equal installs
Release notes excerpt
Fixes vc-update skipping adaptive content migration when the installed version already matches the remote. install.sh writes .vc-version to the new version but cannot run the adaptive legacy-layout migration; vc-update would then bail with 'Already up to date', stranding legacy reports/references dirs and flat plans. vc-update now runs a legacy-artifact scan on version equality and continues to Part D migration when legacy artifacts exist.View on GitHub ↗