Skip to content
/

Clavix

clavix · ClavixDev/Clavix · ★ 656 · last commit 2026-02-18

Enforces strict PRD/implementation phase separation via mandatory State Assertions and a 6-type Self-Correction Protocol across 20+ AI coding tools.

Best whenThe Self-Correction Protocol (Detect-Stop-Correct-Resume loop with 6 named fault types) makes workflow deviations visible and correctable rather than silent.
Skip ifWriting code during PRD phase, Skipping strategic questions
vs seeds
spec-kit(npm install, mirrored commands+skills, full SDLC, multi-tool) but adds the Self-Correction Protocol (6-type fault detec…
Primitive shape 14 total
Commands 4 Skills 10
00

Summary

Clavix — Summary

Clavix is an npm CLI tool (clavix init) that installs a family of 10 agentskills.io-compatible skills into any project, covering the full PRD-to-verification development loop. Its core value proposition is "agentic prompt workflows as skills" — each skill is a self-correcting markdown template that enforces a specific workflow phase, with explicit Self-Correction Protocols that instruct the AI to detect and halt six categories of deviation (writing code during PRD phase, skipping questions, missing sections, etc.) rather than continuing silently. The PRD skill (clavix-prd) is notable for its State Assertion requirement — the AI must output a visible "CLAVIX MODE: PRD Development / Implementation: BLOCKED" banner before starting, ensuring planning and implementation phases are never conflated. At version 7.3.0 with 656 stars, Clavix supports 20+ AI tools across Claude Code, Gemini CLI, Cursor, Copilot, Roocode, and others. It is the most cross-tool-portable PRD framework in this batch. Compared to seeds, it most closely resembles spec-kit (mirrored command/skill pairs, multi-tool install, full SDLC coverage) but adds the Self-Correction Protocol pattern and explicit mode state assertion, which no seed implements.

01

Overview

Clavix — Overview

Origin

Created by ClavixDev (ClavixDev/Clavix). Apache-2.0 license. Version 7.3.0 as of analysis date. Active development (pushed 2026-02-18). TypeScript CLI.

Philosophy

Clavix is built on two premises:

  1. Separation of phases: PRD development and implementation must be explicitly blocked from each other. A skill in "PRD mode" must refuse to write code; a skill in "implement mode" must refuse to write new specs. The State Assertion pattern enforces this visually.

  2. Self-correcting workflows: Rather than relying on the model to stay on track, each skill contains a "Self-Correction Protocol" — a table of 6 mistake types that the AI must detect, halt on, announce explicitly ("I apologize - I was [describe mistake]. Let me return to PRD development."), and correct. This makes deviations visible rather than silent.

From README:

"No code executes during skill invocation. The markdown templates ARE the product." "The agent follows the workflow — using its native tools. Outputs saved locally — under .clavix/outputs/."

Target audience

Developers across multiple tool ecosystems (Claude Code, Cursor, Copilot, Gemini CLI, Roocode, Cline, etc.) who want a portable, self-enforcing PRD-to-implementation workflow without learning a tool-specific framework.

Workflow families

  • Quick: clavix-improve → clavix-implement
  • Full: clavix-prd → clavix-plan → clavix-implement → clavix-verify
  • Exploratory: clavix-start → clavix-summarize → clavix-plan

Manifesto-style quotes

From clavix-prd SKILL.md:

"State Assertion (REQUIRED): Before starting PRD development, output: 'CLAVIX MODE: PRD Development / Mode: planning / Implementation: BLOCKED — I will develop requirements, not implement the feature'"

02

Architecture

Clavix — Architecture

Distribution

  • npm package: clavix
  • Binary: clavix (bin/clavix.js)
  • CLI entry: clavix init → copies skills into project or global scope

Required runtime

  • Node.js >= 18.0.0

Install modes

  • Global: clavix init → installs to ~/.config/agents/skills/
  • Project: clavix init → installs to .skills/ or tool-specific dir

Directory tree (source)

ClavixDev/Clavix/
├── bin/clavix.js
├── src/
│   ├── cli/           # Commander.js CLI entry
│   ├── core/          # Install, update, diagnose logic
│   ├── templates/
│   │   ├── skills/    # 10 skill markdown templates
│   │   │   ├── prd.md
│   │   │   ├── plan.md
│   │   │   ├── implement.md
│   │   │   ├── verify.md
│   │   │   ├── review.md
│   │   │   ├── improve.md
│   │   │   ├── start.md
│   │   │   ├── summarize.md
│   │   │   ├── refine.md
│   │   │   ├── archive.md
│   │   │   └── using-clavix.md
│   │   ├── prd-questions.md        # 5-question PRD flow
│   │   ├── full-prd-template.hbs   # Full PRD Handlebars template
│   │   ├── quick-prd-template.hbs  # Quick PRD template
│   │   ├── agents/                 # Agent-specific integration files
│   │   ├── slash-commands/         # Tool-specific slash command files
│   │   └── instructions/           # AGENTS.md, OCTO.md, etc.
│   ├── config/
│   ├── utils/
│   └── types/
└── schemas/

Output structure (installed)

.clavix/
└── outputs/           # Per-session artifacts (PRDs, plans, etc.)

Target AI tools (20+)

CLI agents: Claude Code, Gemini CLI, Qwen, Droid, CodeBuddy, OpenCode, LLXPRT, Amp, Crush, Codex, Augment, Vibe

IDE extensions: Cursor, Windsurf, Kilocode, Roocode, Cline, GitHub Copilot

Universal: AGENTS.md, OCTO.md, WARP.md

Config files

None required. clavix init writes the skills to the appropriate tool-specific directory based on user selection.

03

Components

Clavix — Components

Skills (10)

Name Purpose
clavix-prd Create PRD through strategic one-at-a-time questions + Self-Correction Protocol
clavix-plan Create task breakdown from PRD with explicit implementation tasks
clavix-implement Execute tasks with progress tracking; implementation BLOCKED until invoked
clavix-verify Verify implementation against requirements
clavix-review Review PRs with criteria presets
clavix-improve Optimize prompts with 6-dimension quality assessment
clavix-start Begin conversational exploration of an idea
clavix-summarize Extract structured requirements from a conversation
clavix-refine Update existing PRD or prompt
clavix-archive Archive completed projects

Templates

File Purpose
templates/prd-questions.md 5-question PRD flow (with validation gates per question)
templates/full-prd-template.hbs Full PRD Handlebars template (11 sections)
templates/quick-prd-template.hbs Quick PRD template (5 sections, for AI consumption)

CLI Commands

Command Purpose
clavix init Initialize or reconfigure integrations
clavix update Regenerate templates
clavix diagnose Check installation health
clavix version Show version

Integration Files

Clavix generates tool-specific integration files based on user selection during clavix init:

  • AGENTS.md — universal agent entry
  • .claude/commands/clavix-*.md — Claude Code slash commands
  • .cursor/rules/ — Cursor rules
  • Tool-specific formats for Windsurf, Copilot, Roocode, etc.

PRD Questions (from prd-questions.md)

  1. "What are we building and why?" (required, minLength:20)
  2. "What are the must-have core features?" (required, minLength:30)
  3. "Tech stack and requirements?" (optional)
  4. "What is explicitly OUT of scope?" (required, minLength:15)
  5. "Any additional context or requirements?" (optional)
05

Prompts

Clavix — Prompts

Excerpt 1: clavix-prd — State Assertion (REQUIRED)

Source: src/templates/skills/prd.md

## State Assertion (REQUIRED)

**Before starting PRD development, output:**

CLAVIX MODE: PRD Development Mode: planning Purpose: Guiding strategic questions to create comprehensive PRD documents Implementation: BLOCKED - I will develop requirements, not implement the feature


Technique: Mandatory mode banner. The AI must output a visible state declaration before any PRD work begins, making the current workflow phase explicit and blocking implementation. This is enforced by the skill description stating it is REQUIRED — there is no code check, just a strong instruction.


Excerpt 2: clavix-prd — Self-Correction Protocol

Source: src/templates/skills/prd.md

## Self-Correction Protocol

**DETECT**: If you find yourself doing any of these 6 mistake types:

| Type | What It Looks Like |
|------|--------------------|
| 1. Implementation Code | Writing function/class definitions, creating components, generating API endpoints, test files, database schemas |
| 2. Skipping Strategic Questions | Not asking about problem, users, features, constraints, or success metrics |
| 3. Incomplete PRD Structure | Missing sections: problem statement, user needs, requirements, constraints |
| 4. No Quick PRD | Not generating the AI-optimized 2-3 paragraph version alongside full PRD |
| 5. Missing Task Breakdown Offer | Not offering to generate tasks.md with actionable implementation tasks |
| 6. Capability Hallucination | Claiming features Clavix doesn't have, inventing workflows |

**STOP**: Immediately halt the incorrect action

**CORRECT**: Output:
"I apologize - I was [describe mistake]. Let me return to PRD development."

**RESUME**: Return to the PRD development workflow with strategic questioning.

Technique: Detect-Stop-Correct-Resume (DSCR) pattern. Unlike superpowers' Iron Law (prohibition statement + rationalization table), Clavix's Self-Correction Protocol is a fault-handling loop: detect the deviation type, halt, announce, resume. The 6-type enumeration is exhaustive and specific, reducing ambiguity about what counts as a violation.


Excerpt 3: PRD Question Flow — Validation Gates

Source: src/templates/prd-questions.md

## Question 1

**text:** 🎯 What are we building and why? (Problem + goal in 2-3 sentences)
**type:** text
**required:** true
**validation:** minLength:20

## Question 2

**text:** ⚡ What are the must-have core features? (List 3-5 critical features)
**type:** text
**required:** true
**validation:** minLength:30

## Question 4

**text:** 🚫 What is explicitly OUT of scope? (What are we NOT building?)
**type:** text
**required:** true
**validation:** minLength:15

Technique: Structured Q&A with validation constraints. Questions 1, 2, and 4 are required with minimum length validation, enforcing substantive answers before proceeding. Questions 3 and 5 are optional with skip instructions, reducing friction for teams with less context.

09

Uniqueness

Clavix — Uniqueness

Differs from Seeds

Closest seed is spec-kit (mirrored skill/command pairs, multi-tool installation, full SDLC coverage, npm install). Clavix diverges in two structural ways: (1) the Self-Correction Protocol — a per-skill fault table instructing the AI to detect, announce, and recover from 6 specific deviation types rather than relying on model compliance; and (2) the State Assertion requirement — each skill must output an explicit mode banner before beginning work, blocking cross-phase contamination visually. Compared to superpowers, Clavix's enforcement is conversational (DSCR loop) rather than pre-emptive (Iron Law rationalization table). Compared to openspec, Clavix is broader in tool support (20+ vs openspec's handful) and lacks openspec's delta-diff spec format. Among seeds, Clavix is the most "cross-tool" framework — explicitly targeting 20+ AI clients via clavix init tool selection.

Positioning

Clavix occupies the "portable, self-enforcing PRD workflow" niche: it works across more AI tools than any other framework in this batch, and its self-correction pattern makes PRD/implementation phase boundaries explicit without requiring Claude Code-specific hooks.

Observable Failure Modes

  1. Self-Correction is advisory only: The DSCR protocol instructs the model to detect deviations but cannot enforce it programmatically. A non-compliant model pass silently ignores the protocol.
  2. Version drift: At v7.3.0, the skill templates shipped via npm may lag behind the source in the repo. clavix update is needed to sync.
  3. No inter-skill orchestration: There is no automated chain — users must manually invoke skills in sequence. Skipping clavix-verify after clavix-implement is trivially easy.
  4. Handlebars templates not rendered by model: The .hbs PRD templates are rendered by the CLI during clavix init to produce static markdown; the model never sees the template syntax. This means the CLI must be re-run to update template content, not just editing the markdown.
04

Workflow

Clavix — Workflow

Three Workflow Families

Full Workflow

clavix-prd → clavix-plan → clavix-implement → clavix-verify

Quick Workflow

clavix-improve → clavix-implement

Exploratory Workflow

clavix-start → clavix-summarize → clavix-plan

PRD Phase (clavix-prd)

Step Action Gate
1 State Assertion output (REQUIRED) Must output mode banner before proceeding
2 Ask Q1: "What are we building and why?" Validate: must have both problem AND goal
3 Ask Q2: core features Validate: minLength:30
4 Ask Q3: tech stack (optional)
5 Ask Q4: out of scope Validate: minLength:15
6 Ask Q5: additional context (optional)
7 Generate Full PRD (full-prd-template.hbs) User review
8 Generate Quick PRD (AI-optimized 2-3 paragraphs)
9 Offer task breakdown → clavix-plan

Self-Correction Protocol (all skills)

Before each action, the skill checks for 6 mistake types:

Type What It Looks Like
1 Writing function/class definitions, creating components
2 Skipping strategic questions
3 Missing PRD sections (problem statement, user needs, requirements, constraints)
4 No quick PRD alongside full PRD
5 Not offering task breakdown at end
6 Capability hallucination (claiming features Clavix doesn't have)

On detection:

  1. STOP immediately
  2. Output: "I apologize - I was [describe mistake]. Let me return to PRD development."
  3. RESUME from the correct step

Artifacts

Phase Artifact
clavix-prd Full PRD, Quick PRD, task breakdown offer
clavix-plan tasks.md with atomic tasks
clavix-implement Code changes + progress report
clavix-verify Verification report vs PRD
clavix-review PR review with criteria presets
All .clavix/outputs/ directory
06

Memory Context

Clavix — Memory & Context

State Storage

File-based under .clavix/outputs/:

  • PRD documents
  • Task breakdowns
  • Verification reports

Persistence

Project-level. Outputs survive session boundaries via the .clavix/outputs/ directory.

Cross-Session Handoff

Yes — the clavix-refine skill is explicitly designed to update an existing PRD, implying cross-session continuation. The clavix-plan skill reads the PRD to generate tasks, completing the chain.

Context Compaction

No explicit handling. The quick PRD (2-3 paragraph AI-optimized version) serves as a lightweight context summary that downstream skills (clavix-plan, clavix-implement) can read without loading the full PRD.

Memory Type

File-based markdown. The .clavix/outputs/ directory is the persistent state.

Search Mechanism

None.

Prompt Chaining

Strong prompt chaining pattern: clavix-prd outputs → clavix-plan reads → clavix-implement reads → clavix-verify reads. Each phase's output is the next phase's primary input.

07

Orchestration

Clavix — Orchestration

Multi-Agent

No. Single agent per skill invocation. Skills chain output artifacts (not agents).

Orchestration Pattern

Sequential. User invokes skills in workflow order. No automatic orchestration between skills.

Isolation Mechanism

None. All writes go to .clavix/outputs/ in the working directory.

Multi-Model

No. Single model.

Execution Mode

Interactive-loop. Each skill requires user interaction (Q&A, confirmation).

Crash Recovery

Partial — file outputs persist. The clavix-refine skill allows resuming a PRD update. No explicit state machine.

Context Compaction

The Quick PRD artifact (2-3 paragraphs) functions as a compressed context summary for downstream skills.

Cross-Session Handoff

Yes — .clavix/outputs/ persists across sessions.

Prompt Chaining

Yes — explicit chain: prd → plan → implement → verify. Each skill's output file is the next skill's input.

Subagent Definition Format

None.

08

Ui Cli Surface

Clavix — UI & CLI Surface

CLI Binary

  • Name: clavix
  • Package: clavix (npm)
  • Entry: bin/clavix.js
  • Version: 7.3.0

Subcommands

Command Purpose
clavix init Initialize or reconfigure integrations; copies skills to tool-specific dirs
clavix update Regenerate templates with latest version
clavix diagnose Check installation health
clavix version Show version

Local UI

None. No dashboard.

IDE Integration

Broadest multi-tool integration in this batch:

Category Tools
IDEs Cursor, Windsurf, Kilocode, Roocode, Cline, GitHub Copilot
CLI agents Claude Code, Gemini CLI, Qwen, Droid, CodeBuddy, OpenCode, LLXPRT, Amp, Crush, Codex, Augment, Vibe
Universal AGENTS.md, OCTO.md, WARP.md

Command format varies by tool:

  • CLI tools (Claude Code, Gemini): /clavix:improve
  • IDE extensions (Cursor, Copilot): /clavix-improve

Observability

No audit log. .clavix/outputs/ contains produced artifacts but no structured execution log.

Related frameworks

same archetype · same primary tool · same memory type

BMAD-METHOD ★ 48k

Provides a full agile delivery lifecycle with named expert-persona AI collaborators that elicit the human's best thinking rather…

Agent OS ★ 4.6k

Extracts implicit codebase conventions into token-efficient markdown standards files and injects them selectively into AI agent…

Claude Conductor ★ 367

Gives Claude Code a persistent, cross-linked, auto-analyzed documentation system so it retains codebase context across sessions.

Spec-Driver (Greenfield Spec-Driven Development) ★ 25

Prevents spec rot in AI-assisted development by making implementation changes flow back into evergreen, authoritative specs via…

Anthropic Knowledge Work Plugins ★ 16k

Role-specialized plugin bundles with live MCP connectors that turn Claude into a domain expert for enterprise knowledge workers.

Codex Integration for Claude Code (skill-codex) ★ 1.3k

Single Claude Code skill that handles Codex CLI invocation correctly (stdin blocking, thinking token suppression, session resume)…