Spec-Driven Development with Coding Agents
Updated 22 May 2026
Introduction
https://www.deeplearning.ai/courses/spec-driven-development-with-coding-agents
Why spec-ddriven development?
Benefits of spec-driven developmetn with coding agents
- Control code with small changes to spec
- Eliminate context deay
- Improve intent fidelity
agent is muscle, spec is the brain
Workflow overview
Need the right level of detail
consitution: mission (why), tech stack (engineering team), roadmap (living document)
feature phase: specifiction -> implementation -> validation
- you: design, superise, review, accept, or ask for chagnes
Set up your envirornment
VSCode + Claude Code
Webstorm is jetbrain
Setup
Creating the constitution
consitution mission - what’s the core idea? tech stack - how does it fit wit hte tech stack raodmap - what feature are planned?
Ask agent to make changes, so agent has context.
specs/mission.md'
specs/tech-stack.md’
`specs/roadmap.md’
Feature specification
Start a new session
specs/plan.md'
specs/requirements.md’
`specs/validation.md’
Feature implementation
Feature validation
human in the loop. Need changes to be manageable to reduce cognitive debt.
Dont’ do stuff manually since it will cause things to get out of sync.
Project replanning
Tell agent we want responive design since mobile. Small frequent commits.
skill for changelog update. skills for writing skills.
repeatble process should be skills
The second feature phase
have clean division between each feature phase to cope with ai fatigue
Stick to higher level requirements. dont’ care about variable names.
Use subagetns to do deep reviews.
The MVP
Implement rest of the roadmap when you feel confidence in feature spec and constitution.
Legacy support
Greenfield vs brownfield, new code versus existing code.
Build your own workflow
Back to agent skills. Use skill creator. skills per project or global. need to reload claude to know skills, skills can be called from other skills. name a skill like @ if you want to use it. agents moving from custom slash commands to skills.
context7, mcp for information about packages. cli + skills recommended over mcp.
acclerating chagne between mcp -> cli + skills, less tokens and context usage.
claude plugins
github spec-kit
opensec from fissionai
keep backlog file for later research.
Agent replaceability
mcp, agents.md, agent skills, acp - connect agents to clients.
acp registry, open code is in acp registry.
acp includes next edit suggestion