Agent Specifications Guide

Generate context-rich specifications for AI coding agents from structured product thinking

What Are Agent Specifications?

Agent specifications are structured, context-rich specifications designed for AI coding agents like GitHub Copilot, Claude Code, and Devin. Unlike traditional PRDs or user stories, agent specs include your product vision, strategic guardrails, acceptance criteria, success metrics, and learnings from previous implementations.

Delvyn Studio assembles all of your product thinking into specifications that AI agents can implement correctly on the first try. Follow the 8 steps below to generate your first agent specification.

1
Connect Your Integration

Connect GitHub, Linear, or Jira in Settings > Integrations. This is where your agent specifications will be pushed.

  • •GitHub: Specs are pushed as Issues that GitHub Copilot Agent can pick up
  • •Linear: Specs become tasks that Claude Code, GitHub Copilot, or Devin can implement
  • •Jira: Specs are created as stories with full context attached
2
Define Your Product Vision

Go to your Product > Vision page and define your product vision. This becomes the strategic context that every agent specification inherits.

  • •Set your product's purpose and target users
  • •Define strategic pillars that guide what to build
  • •AI can help generate your initial vision from a brief description
  • •The vision flows into every spec as strategic guardrails
3
Create a Product Idea

Navigate to Product > Ideas and create an idea. Ideas capture the 'what' and 'why' before you commit to building.

  • •Describe the feature or improvement
  • •Link it to a strategic pillar for alignment
  • •Add discovery research if you've validated the idea
  • •AI suggests potential risks and considerations
4
Set OKR Key Results

Create or select an OKR cycle, add an objective, and define key results with success criteria. These become the measurable outcomes in your agent spec.

  • •Key Results define what success looks like
  • •Success criteria become acceptance tests in the spec
  • •AI helps suggest measurable key results
  • •Link key results to product ideas for traceability
5
Generate the Agent Specification

From the idea or key result, click 'Generate Agent Spec.' The engine assembles your product vision, strategy, discovery insights, OKR criteria, and learnings into a structured specification.

  • •Vision and strategic guardrails are included automatically
  • •Acceptance criteria come from your OKR success criteria
  • •Discovery insights provide customer context
  • •Previous learnings improve spec quality over time
6
Review with AI Coach

Before pushing, the AI Specification Coach reviews your spec for completeness, ambiguity, and potential implementation issues.

  • •Coach checks for missing context or acceptance criteria
  • •Flags ambiguous requirements that could lead to wrong implementation
  • •Suggests improvements based on learnings from previous specs
  • •You can accept, modify, or dismiss suggestions
7
Push to Your Integration

Push the reviewed spec to GitHub, Linear, or Jira. The AI coding agent in your team's workflow picks it up and implements it.

  • •GitHub: Creates an Issue with full context for Copilot Agent
  • •Linear: Creates a task with embedded specs for Claude Code / GitHub Copilot
  • •Context Sync: Optionally push spec files to your repository
  • •The spec includes everything the agent needs to implement correctly
8
Capture Learnings

After the code ships, capture what worked and what didn't. These learnings feed back into future specs, creating a continuous improvement loop.

  • •Record what the agent implemented correctly
  • •Note any missing context that caused issues
  • •Flag patterns that improve or degrade spec quality
  • •The Spec Quality Dashboard tracks trends over time
Integration-Specific Workflows

GitHub + Copilot Agent

  1. 1.Connect GitHub in Settings > Integrations
  2. 2.Generate and review your agent spec
  3. 3.Push to GitHub as an Issue
  4. 4.Assign to Copilot Agent for implementation
  5. 5.Copilot creates a PR with full context

Linear + Claude Code / GitHub Copilot

  1. 1.Connect Linear in Settings > Integrations
  2. 2.Generate and review your agent spec
  3. 3.Push to Linear as an agent task
  4. 4.Claude Code or GitHub Copilot picks up the enriched task
  5. 5.Implementation includes full product context

Ready to generate your first spec?

Start by defining your product vision, then create an idea and generate a spec.

Get Started