Spec-Driven Agentic Development — Summary

Spec-Driven Agentic Development by marcelsud is a minimal Claude Code command pack (5 slash commands) that installs via npx degit into any project's .claude/ directory. Its defining characteristic is the use of EARS (Easy Approach to Requirements Syntax) formatted requirements as the spec standard and a TDD-first task structure. The spec/create command spawns multiple parallel Task agents (Explore agents for codebase analysis, Plan agents for strategy) before generating the three canonical artifacts: context.md, requirements.md, and tasks.md — all stored under features/<feature-name>/. It is the most direct derivative of the kiro requirements/design/tasks pattern in this batch, but implemented purely as Claude Code slash commands with no CLI binary, no hooks, no MCP server. With 172 stars and 1 contributor, it is a focused personal tool made public. Compared to seeds, it is closest to openspec in its 1:1 command-to-workflow-phase mapping but uses EARS format (borrowed from kiro) rather than open-ended requirements.

Spec-Driven Agentic Development — Overview

Origin

Created by Marcelo Santos (@marcelsud, LinkedIn: marcelsud). Published as a GitHub repository for community use. No npm/pip package — installed via npx degit. 172 stars, 1 contributor, last active December 2025.

Philosophy

A structured methodology for building software features through specifications, optimized for AI-powered development tools. The core workflow is:

/spec:create [feature]  →  context.md + requirements.md + tasks.md
/spec:execute [feature] →  Implementation

The framework's design choices:

1. EARS-formatted requirements — uses Easy Approach to Requirements Syntax for unambiguous requirement statements
2. TDD task breakdown — tasks.md enforces test-driven development steps
3. Multi-agent exploration — spec creation uses parallel Task agents to explore codebase before generating specs
4. Minimal footprint — 5 commands, installs into .claude/commands/spec/

EARS Format

Requirements use EARS notation (same as kiro). EARS provides structured, unambiguous requirement templates:

• "When [condition], the system shall [behavior]"
• "While [condition], the system shall [behavior]"
• "The [feature] shall [behavior]"

This prevents vague requirements that are ambiguous to implement.

No Manifesto

The README is minimal — no philosophical positioning, no competitive comparisons. Pure implementation.

Spec-Driven Agentic Development — Architecture

Distribution

• Type: Claude Code command pack (standalone-repo)
• Install: npx degit marcelsud/spec-driven-agentic-development/.claude .claude
• Language: Markdown (command files only)
• Version analyzed: latest main (December 2025)

Installation

cd /path/to/your/project
npx degit marcelsud/spec-driven-agentic-development/.claude .claude

This copies the .claude/ directory (CLAUDE.md + commands/spec/) into the project. No package manager, no binary.

Directory Tree (post-install)

your-project/
├── .claude/
│   ├── CLAUDE.md               # Methodology reference
│   └── commands/spec/
│       ├── create.md           # Feature spec creation
│       ├── execute.md          # Implementation execution
│       ├── list.md             # List features
│       ├── status.md           # Show implementation status
│       ├── help.md             # Command help
│       ├── diagram.md          # Generate diagrams
│       ├── next.md             # Next action
│       └── review.md           # Review feature
└── features/
    └── [feature-name]/
        ├── context.md          # Context + technical decisions
        ├── requirements.md     # EARS-formatted requirements
        ├── exploration.md      # Codebase exploration findings
        └── tasks.md            # TDD task breakdown

Required Runtime

• None (beyond having Claude Code installed)
• No Python, Node.js, or other runtime required
• Git (optional, for the project itself)

Target AI Tools

Claude Code exclusively. The degit install copies Claude Code slash commands.

Spec-Driven Agentic Development — Components

Commands (8 slash commands)

Skills

None.

Subagents

Yes — the /spec:create command spawns parallel Task agents:

• Explore agents (1-3, based on complexity): Analyze codebase patterns, find similar features, analyze tech stack
• Plan agents: Generate implementation approaches and strategies

Spawned via Claude Code's Task tool with subagent_type: "Explore". This is the framework's most distinctive engineering choice.

Hooks

None. No .claude/settings.json.

MCP Servers

None.

Scripts

None.

Templates

3 artifact templates per feature (generated by the commands):

• context.md — context and technical decisions
• requirements.md — EARS-formatted requirements
• tasks.md — TDD task breakdown

State Files

• features/<feature-name>/context.md
• features/<feature-name>/requirements.md
• features/<feature-name>/exploration.md
• features/<feature-name>/tasks.md

CLAUDE.md

Includes a methodology reference document that loads the spec-driven workflow context into every Claude session.

Spec-Driven Agentic Development — Uniqueness & Positioning

differs_from_seeds

Closest to openspec in structure (command-per-phase, each command produces one artifact) but uses EARS format borrowed from kiro rather than open-ended requirements. The key differentiator from all seeds and all other frameworks in this batch is the complexity-triage-gated parallel Explore agent spawning: before generating any spec artifact, the framework assesses complexity and dispatches 1-3 parallel Task subagents to explore the codebase, producing an exploration.md that grounds the spec in real code patterns. No seed framework does this — superpowers' subagent dispatch is for implementation, not spec generation. This makes spec-driven-agentic-development the only framework where spec quality directly scales with codebase complexity assessment.

Positioning

Minimal Claude Code command pack for developers who want spec-driven development without installing a Python CLI or npm package. The degit install pattern is zero-friction.

Distinctive Opinion

Spec quality requires codebase exploration first. The exploration phase is mandatory (not optional) and automatically scales (1-3 agents based on complexity). This prevents specs from being written in isolation from the actual codebase they'll modify.

Observable Failure Modes

1. Claude Code only: No multi-runtime support.
2. No hooks: Nothing enforces the workflow sequence — agents can skip phases.
3. No task state tracking: tasks.md has no progress tracking (unlike specs-workflow-mcp's .workflow-confirmations.json).
4. 1 contributor: Maintenance risk.

Explicit Antipatterns

None documented.

Cross-References

EARS format borrowed from kiro. Requirements/design/tasks artifact structure also matches kiro pattern.