VIBE Framework
Challenge
AI-generated code suffers from systematic problems: hardcoded API keys, tests that only pass with mocks, OWASP vulnerabilities, infinite retry loops on the same error. Developers correct the same issues session after session because AI doesn't remember.
Approach
- 12 skills (8 domain + 4 utility): security (Heimdall), testing with 8 personas (Emmet), UI (Seurat), SEO+GEO (Ghostwriter), CRO (Baptist), video (Orson), documents (Scribe), meta-skill (Forge), setup, reflect, pause, resume
- 6 deterministic mechanical hooks — not suggestions AI can ignore, but gates that block on violation
- 9 agents (2 general-purpose + 7 domain audit) with separate context and persistent memory
- Market intelligence: shared competitor research protocol across Ghostwriter, Seurat, and Baptist in 11 languages
- Audit orchestrator (/vibe:audit) with delta analysis and automatic regression detection
- Self-learning system: captures corrections in 6 languages, saves to project or user memory
- 59 automated tests covering plugin structure, skills, hooks, agents, and security patterns
Results
Open source framework released as a Claude Code plugin, now at v3. Solves the most common problems in vibe-coded software — security, quality, inter-session amnesia — and adds market intelligence for decisions informed by real-world competitive data.
Claude Code out-of-the-box optimizes for speed and token savings. VIBE inverts this: quality above all, even at the cost of more tokens and longer execution.
With v3, VIBE introduces market intelligence as a core principle: skills discover answers through global competitor research rather than asking unanswerable questions. Three principles guide the framework: market intelligence over guesswork, process discipline over knowledge, and mechanical quality gates.
The framework operates on three layers. Skills encode expert methodologies — complete workflows with numbered phases, verification steps, and reference material loaded on-demand. Hooks are deterministic mechanical gates: a shell script scanning for hardcoded API keys can't be persuaded to skip the check. Agents operate in separate contexts with persistent memory — the reviewer hasn't seen the implementation, so it can't exhibit self-review bias.
VIBE includes 12 skills (8 domain + 4 utility), 6 hooks that block on failure, and 9 agents (2 general-purpose + 7 domain audit). Ghostwriter, Seurat, and Baptist share a competitor research protocol across 11 languages, with cached results and automatic 30-day refresh. The system learns from user corrections and applies them in subsequent sessions.