specprompt.com
Spec  ·  How  ·  GitHub
Spec-Driven Development

SpecPrompt

Prompts are transient. Specs persist.
Define agent behavior as versioned, testable specifications — not disposable chat messages. The spec becomes the source of truth that survives tool-switching, team changes, and model upgrades.

"I see spec-driven development as the future maturation of software development in the agentic age." — Brad Shimmin, Futurum Group (Feb 2026)

From Vibe Coding to Spec-Driven Development

In 2025, "vibe coding" — typing a few prompts and letting AI generate code — went mainstream. By late 2025, Thoughtworks, GitHub, AWS (Kiro), and InfoQ were all pointing to the same conclusion: vibe coding creates today's productivity and tomorrow's technical debt.

Spec-Driven Development (SDD) separates intent from implementation. You write a specification — requirements, acceptance criteria, boundary conditions, architectural constraints — and the AI generates code against it. The spec is the artifact that persists, versioned alongside your codebase. SpecPrompt makes this practical for agent development.

❌ Vibe Coding

# Chat message (lost after session)

"Build me a customer support agent
that can look up orders and process
refunds. Make it friendly."

# Problems:
# - No boundary conditions
# - No acceptance criteria
# - No version control
# - Lost when you switch tools

✓ SpecPrompt

# SPEC.md (versioned, testable)

## Agent: customer-support-v2
## Capabilities:
  - orders:read
  - refunds:create (max $500)
## Constraints:
  - Never disclose internal pricing
  - Escalate refunds > $500 to human
## Tests:
  - Given order #123, return status
  - Reject refund > $500 gracefully

The SpecPrompt Workflow

01 — Define

Write a SPEC.md in plain Markdown with structured sections: capabilities, constraints, acceptance tests, boundary conditions. Human-readable, machine-parseable.

02 — Generate

AI agents (via Agentelic or any SDD-compatible tool) generate implementation from the spec. Code is the output — the spec is the authority.

03 — Test

Acceptance criteria from the spec become automated tests. Every build validates behavior against the contract. Drift is caught before deployment.

04 — Version

Specs live in git alongside code. Branching, diffing, merging — all the version control practices you already know, applied to agent behavior.

05 — Share

Publish specs to FleetPrompt as reusable templates. Teams fork specs, adapt them, and contribute improvements back. A marketplace of proven agent patterns.

06 — Evolve

Specs are living documents. As Graphonomous learns from production, insights feed back into spec refinements. The spec-agent-learn loop closes.

The Industry Is Moving Here

InfoQ (Feb 2026): SDD at Enterprise Scale

"With uninterrupted agent execution increasingly replacing interactive prompting, intent articulation becomes even more critical. Specs become the shared interface where product, architecture, engineering, and quality collaborate."

AWS Kiro: Spec-Driven AI Development

Amazon's new AI IDE centers on spec-driven development — taking prompts, generating requirements in EARS notation, then building architecture and code from those specs. Validates that SDD is the direction the largest cloud provider is betting on.

GitHub Spec Kit (Open Source)

GitHub released Spec Kit as an open-source toolkit for generating requirements, plans, and tasks to guide coding agents through iterative, structured development. Microsoft's AI-led SDLC blog calls it the "entry point that triggers the flow."

Thoughtworks Technology Radar

"The core of SDD goes beyond vibe coding, separating the design and implementation phases. Spec-driven development remains an emerging practice as 2025 draws to a close; we're likely to see even more change in 2026."

SPEC.md — A Living Standard

SpecPrompt defines a Markdown-based specification format designed to be both human-readable and machine-parseable. It's not a new language — it's structured Markdown with conventions that AI agents can reliably interpret.

SPEC.md Structure

---
name: order-fulfillment-agent
version: 2.1.0
runtime: opensentience
author: ops-team
---

## Purpose
Process incoming orders, validate inventory,
trigger fulfillment workflows.

## Capabilities
- inventory:read
- orders:read,write
- fulfillment:trigger
- notifications:send

## Constraints
- Never process orders exceeding $10,000
  without human approval
- Rate limit: max 100 orders/hour
- Retry failed fulfillment max 3 times

## Acceptance Tests
- Given valid order → fulfill within 30s
- Given out-of-stock item → notify customer
- Given order > $10K → escalate to human
- Given fulfillment failure → retry 3x, then alert

## Dependencies
- graphonomous (knowledge graph)
- inventory-api (MCP server)
- notifications-service (MCP server)

What's Coming

Q1 2026

SPEC.md Format v1.0

Open standard for agent specifications. Markdown-based, machine-parseable. Reference parser in Elixir and TypeScript.

Q2 2026

CLI + MCP Server

Validate, lint, and test specs from the command line. MCP server exposes spec operations as tools — any MCP client can create and manage specs.

Q3 2026

Agentelic Integration

SpecPrompt specs become first-class inputs to Agentelic's build pipeline. Write spec → generate agent → test → deploy.

Q4 2026

FleetPrompt Spec Registry

Publish, discover, and fork agent specs on FleetPrompt. Community-reviewed spec templates for common agent patterns.

The Standards Layer

SpecPrompt is the standards layer of the [&] portfolio. Every product in the ecosystem consumes or produces specs — it's the shared language that makes the stack interoperable.

Integration Map

SpecPrompt defines the contract → Agentelic implements and tests it → OpenSentience enforces permissions from it → Graphonomous learns within its boundaries → FleetPrompt distributes it → Delegatic orchestrates agents described by it.