The agentic development standard to mitigate cognitive technical debt.
🌍Repositorio de VibeSpec en Español
VibeSpec is the open standard for collaboration between humans and artificial intelligence systems. Born in 2026, when producing with AI no longer consists of writing the result by hand but of directing intent, and AI agents have become teammates.
It was born in software development — it remains its flagship use case and the most mature within the standard — but its principles do not depend on code: they apply to any discipline where a person directs an AI agent and needs to retain control over the quality and security of the outcome (writing, translation, data analysis, legal documents, design, research, etc.). What changes across disciplines is not the method, but the concrete verification gate applied in each — see Skills for Hermes Agent below for the detail of this design.
It defines principles, roles, checklists, and metrics that enable any team—regardless of size, discipline, experience level, or chosen tool—to:
- Translate ideas into executable specifications (Spec‑Driven Development)
- Maintain human control over quality and security
- Learn to collaborate with AI agents progressively
- Scale from a rapid prototype to a mission-critical system
| Current Situation | With VibeSpec |
|---|---|
| Each tool has its own vocabulary and workflow | Unified glossary that maps common concepts across all platforms |
| "Vibe coding" produces unreviewed deliverables that reach production | Mandatory checklists that prevent risks without stifling creativity |
| The person's role facing AI is unclear | New professional role as architect, auditor, and decision-maker |
| Lack of traceability in AI decisions | Complete traceability from intent to final outcome |
| Steep and disorganized learning curve | Lite → Pro adoption path with maturity metrics |
The concrete checklist and metric examples in this table are developed today for software development (the most mature use case of the standard); the mechanism transfers equally to other disciplines as their own gates are defined.
vibespec\
├── checklists\
│ ├── templates\
│ │ ├── example-deviation.md
│ │ ├── template-markdown.md
│ │ └── template-yaml.yaml
│ ├── README.md
│ └── vibespec-check.md
├── governance\
│ ├── adoption.md
│ ├── deviations.md
│ ├── evolution.md
│ └── README.md
├── skills\
│ ├── README.md
│ ├── vibespec-lite.md
│ └── vibespec-pro.md
├── hermes-skills\
│ └── vibespec\
│ ├── README.md
│ ├── vibespec-lite\
│ │ └── SKILL.md
│ └── vibespec-pro\
│ └── SKILL.md
├── standard\
│ ├── glossary.md
│ ├── prompt-templates.md
│ ├── vibespec-lite.md
│ └── vibespec-pro.md
├── LICENSE.md
├── README.md
└── TRADEMARK.md
-
Choose your level
- If you're a beginner or working on a small project →
standard/vibespec-lite.md - If you work in an enterprise or high-criticality environment →
standard/vibespec-pro.md
- If you're a beginner or working on a small project →
-
Download the master checklist
checklists/vibespec-check.mdand paste it into your next deliverable (PR, draft, report...). -
Create your context file Add an
AGENTS.mdor.cursor/ruleswith the instructions the AI must follow. -
Install the VibeSpec Skill in your agent (Cursor, Windsurf, Claude Code) to automate reminders.
-
Use the Hermes Skills (see below) if your agent is Hermes Agent, to load the standard as a native skill.
In addition to the system prompts for other IDEs, VibeSpec ships native
Hermes Agent in
hermes-skills/vibespec/:
| Skill | Level | What it does |
|---|---|---|
vibespec-lite |
Lite | Applies the C1‑C6 gate and requires human verification before commit. |
vibespec-pro |
Pro | Applies phase verification (P/R/F/D/DP) by reading the master checklist. |
See hermes-skills/vibespec/README.md for
the design (one skill parameterized by discipline, not one per field). Today
both skills read the single checklist at checklists/vibespec-check.md,
centered on software development; the design leaves room to add checklists for
other disciplines without creating new skills.
Install in your Hermes profile (local):
# Copy the skill folder into your profile
cp -r hermes-skills/vibespec/vibespec-lite \
~/.hermes/profiles/<your-profile>/skills/vibespec/
cp -r hermes-skills/vibespec/vibespec-pro \
~/.hermes/profiles/<your-profile>/skills/vibespec/Or, from a Hermes session, ask it to load the skill with
skill_view(name="vibespec-lite"). The agent will read the canonical
checklist from checklists/vibespec-check.md instead of hardcoding it, so the
skill never drifts when you update the standard.
- Explicit intent before execution
- Mandatory human verification in every cycle
- Context as a versionable project asset
- Traceability and change control for all decisions
- Security by design (Secure‑by‑Design)
- Human governance of architectural decisions
- Continuous improvement of the process itself
These seven principles are deliberately discipline-agnostic: they mention no code, text, or concrete artifact. What is specific to each field are the gate checklists that implement them (today, the software development one).
Whoever directs the agent is no longer a mere task executor: they are the context architect, the security and stability auditor, the quality curator, and the final decision-maker. The AI proposes, but the responsibility — technical, legal, and ethical — remains human.
In software development this translates to the programmer ceasing to be a code typist and becoming that architect/auditor; in other disciplines the same role shift applies to the editor who directs an agent's writing, the analyst who audits an AI-generated data pipeline, or any professional supervising an agentic workflow.
| Level | Active Checklist | Expected Compliance |
|---|---|---|
| Lite | C1‑C6 | ≥ 90% of deliverables |
| Structured | Lite + Planning + Functionality | ≥ 85% of deliverables |
| Pro | All phases (including deployment) | ≥ 80% of deliverables |
The unit of measure ("deliverable") replaces the original "PR" here so as not to assume the output is always code; in software development a deliverable is, in practice, still a PR.
VibeSpec is a living standard. If you find an area for improvement, a new risk to cover, want to share a useful template, or want to propose a checklist for a discipline other than software:
- Open an issue describing your proposal.
- Comment and discuss for two weeks (minor changes) or one month (major changes).
- If consensus is reached, it will be integrated in the next revision.
Consult governance/evolution.md for the
complete process.
The standard content is under Creative Commons BY 4.0.
Code snippets are under MIT.
The VibeSpec™ trademark has a usage policy in TRADEMARK.md.
VibeSpec is a free and open standard, created to solve a real problem: bringing order to the chaos of AI-assisted production.
If this standard saves you hours of review, helps you deploy with confidence, or gives you a framework to train your team, consider buying us a coffee via the link below:
Every contribution allows us to maintain, improve, and share VibeSpec with the community.
Designed with the precision of a standard and a passion for clean code.
Created by Franklin A. Lara P. (@FrkL81)