VibeCrafted The Founders' Framework
A convergence-driven system for shipping software with AI agents.
A convergence-driven system for shipping software with AI agents.
The vocabulary is vivid because it was discovered in real founder workflows. Underneath it sits a rigorous operating model for context, variance reduction, and release truth.
LLMs fail when they code blind. By externalizing architectural memory into AST-augmented mappings, VibeCrafted keeps frontier models oriented precisely inside your repository.
Code generation is stochastic. The Marbles loop uses priority-weighted denoising loops to relentlessly reduce variance, forcing the system into a shippable state.
Structured Human-AI Collaborative Engineering. We wrap deterministic human checkpoints around stochastic generation. The result is a system you control and understand.
The substrate that keeps frontier agents anchored to repo truth.
Perception Engine
AST-augmented architectural memory. The eyes for every agent — forces LLMs to see the full repo structure before modifying a single file.
loct.ioHistory Retrieval
Indexes decisions, intents, and outcomes from every AI session. Agents retrieve what was decided and why, ensuring session-to-session continuity.
github.com/VetCoders/ai-contextersAudit & Quality Gate
Generates durable PR review packs with line-level findings. The merge gate that turns terminal dumps into actual artifacts.
github.com/VetCoders/prviewVisual Reviewer
Turns narrated screencasts into structured engineering findings. The orchestrator's eye on the real product surface.
github.com/VetCoders/screenscribeA founder-grade operating model for shaping, denoising, and releasing software with AI agents.
%% VibeCrafted Pipeline v1.0.4
flowchart LR
subgraph "Craft"
scaffold["Scaffold"] --> init["Init"]
init --> workflow["Workflow"]
workflow --> followup["Follow-up"]
end
followup --> marbles{{"Marbles Loop"}}
marbles -->|P0/P1/P2 > 0| marbles
marbles --> dou["DoU"]
subgraph "Ship"
dou --> decorate["Decorate"]
decorate --> hydrate["Hydrate"]
hydrate --> release(["Release"])
end
Scaffold the architecture, init the context, and run the ERi workflow. Force agents to respect the AST and history.
Systematically reduce entropy through Marbles loops until all P0, P1, and P2 issues are resolved.
Audit the product surface with DoU, polish with Decorate, and package with Hydrate before final Release.
Hard-coded procedures born from pain. Powered by frontier models and Rust-based structural intelligence.
Founder-first architecture planning. Maps the landscape and produces a scoped build plan for agents to execute.
Bootstrap the agent session. Forces models to index history (AICX), read the AST (Loctree), and verify quality gates.
The ERi Pipeline: Examine → Research → Implement. Structured implementation with cumulative context feeding.
Spawn external fleet (Codex, Claude, Gemini) in real terminals with isolated context and durable artifacts.
Triple-agent triangulation. Throw all three models at the same problem and synthesize a gap-free research document.
Executive pairing. Shared brain mode for strategic debugging and architectural triage with delegated fieldwork.
Autonomous delivery. Takes a mandate, decides, implements, verifies, and packages without ceremony.
The reality check. Post-implementation audit that gives a brutal GO / NO-GO verdict based on quality gates.
Iterative convergence. Loop until entropy is crushed (P0=0, P1=0, P2=0). Adaptive denoising for code.
PR review pipeline. Generates prview artifacts and squeezes maximum findings with line-level evidence.
Ruthless cleanup. Defines the runtime cone and cuts dead weight, experiments, and scar tissue without mercy.
Safe native delegation using Claude's Task tool. In-session parallelization for small, bounded tasks.
Definition of Undone. Audits product surface, install paths, and SEO. Measures the gap to launch.
Visual coherence. Detects brand tokens and upgrades weak patterns into a premium interaction language.
Market packaging. Executes the non-code work: marketplace listings, SEO fixes, and onboarding flows.
Final milestone. Tags, changelogs, registry publishing, and go-to-market mechanics. Ship it to people.
AI agents are stochastic parrots. Every generation introduces noise. You cannot prompt your way out of it. You iterate.
There is no fixed loop count. The loop runs until P0=P1=P2=0 — the circle is full. Early loops make large structural corrections; late loops make small refinements.
Interactive Playground
The board is no longer squeezed into the landing page. The full-viewport playground is now the teaching surface: manual phases, command prompt, and board-first motion.
Board-first. Manual. Command-shaped. Built for the 1.3.0 direction.
Convergence Trajectory — loop until the circle is full
Score: 15 → 40 → 65 → 85 → 100
AI agents have solved the hardest part of software creation: making the thing work. They have not yet solved the second hardest part: making the thing reachable.
The gap is rarely code quality. It is the space between a working system and a product that people can discover, install, trust, and buy.
The Plague Diagnostic:
Items 4–6 are the Definition of Undone. Phase 3 forces the system into the market.
Bootstrap the control plane, verify the surface, and enter the first founder-grade workflow with the new command deck.
Step 1
Stage the framework locally and let the installer walk through the real environment before it touches the system.
curl -fsSLO https://raw.githubusercontent.com/VetCoders/vibecrafted/main/install.sh
bash install.sh
Step 2
The main front door is now vibecrafted. Daily skill surfaces are exposed as
direct vc-* commands.
vibecrafted help
vc-init claude
Step 3
Once context is anchored, drive the loop deliberately instead of improvising prompt-by-prompt.
vc-marbles codex --count 4 --depth 4
vibecrafted doctor
The bibliography below follows the explicit influence list from VIBECRAFT.md.
Morris, K.; Böckeler, B. (2026). Humans and Agents in Software Engineering Loops. Thoughtworks / Martin Fowler.
Gong, L.; Elhoushi, M.; Cheung, A. (2024). AST-T5: Structure-Aware Pretraining for Code Generation.
Fowler, M. (2009). Technical debt and the later debt literature.
VetCoders (2026). The VibeCraft Manifesto.