DEV Community

Anton Abyzov
Anton Abyzov

Posted on

I Built a Framework That Lets AI Code While I Sleep — Here's What I Learned After 480 Releases

#ai #opensource #devtools #buildinpublic

Six months ago, I got tired of losing context every time a chat session ended.

I was building a sports management platform (14 microservices, Kubernetes, the works) and a soccer training app simultaneously. Every time I started a new Claude or Cursor session, I had to re-explain the architecture, the decisions we'd made, the patterns we'd chosen. Hours of productive context, gone.

So I built SpecWeave.

What is SpecWeave?

SpecWeave is a spec-driven development framework where specifications are the source of truth, not code comments or wiki pages that nobody updates.

npm install -g specweave
Enter fullscreen mode Exit fullscreen mode

The core idea: you describe what you want in increments, and SpecWeave creates persistent documentation that AI agents can reference across sessions.

/sw:increment "User authentication with JWT"
/sw:auto  # Ship while you sleep
Enter fullscreen mode Exit fullscreen mode

What happens under the hood:

.specweave/increments/0001-auth/
|- spec.md    <- WHAT: User stories, acceptance criteria
|- plan.md    <- HOW: Architecture decisions, tech choices
|- tasks.md   <- DO: Implementation tasks with tests
Enter fullscreen mode Exit fullscreen mode

The Problem It Solves

Every AI coding tool has the same fundamental limitation: context dies when the session ends.

You spend 2 hours explaining your architecture to an AI assistant. It writes great code. Session ends. Next session? Starting from scratch.

SpecWeave creates permanent, structured documentation that:

  • Persists across any number of sessions
  • Works with Claude Code, Cursor, Windsurf, Copilot -- any AI assistant
  • Keeps specs, plans, and tasks in sync
  • Tracks decisions with full reasoning

Six months later, you can search "OAuth" and find exactly why you chose that approach, who approved it, and what edge cases you considered.

99% Token Savings with Lazy Loading

One of the features I'm most proud of: SpecWeave only loads plugins when your prompt actually needs them.

Say "React frontend" and the frontend plugin loads. Say "deploy to Kubernetes" and the k8s plugin loads. Everything else stays dormant.

Scenario Without Lazy Loading With Lazy Loading
Non-SpecWeave work ~60k tokens ~500 tokens
SpecWeave work ~60k tokens ~60k (when needed)

This was critical for making it practical in daily use. Nobody wants a framework that burns tokens just sitting there.

What I've Learned After 480 Releases

1. Ship small, ship fast

I publish multiple times a day. Each release is a small improvement. The npm registry shows 483 versions as of this week. Some people think that's excessive. I think it's the only way to iterate fast enough when your users (AI agents) don't file bug reports.

2. Specs > Code

The biggest mental shift: specifications are more valuable than the code they produce. Code can be regenerated from good specs. Specs can't be reverse-engineered from code.

When I started treating specs as first-class artifacts, the quality of AI-generated code improved dramatically. The AI had clear requirements, acceptance criteria, and constraints. It stopped guessing.

3. Autonomous execution is real (with guardrails) 

[08:23:41] [Planning]      Analyzing T-003: Implement refresh token rotation
[08:24:12] [Implementing]  Writing src/auth/token-manager.ts
[08:25:33] [Testing]       Running tests... FAILED
[08:25:47] [Fixing]        Adjusting implementation...
[08:26:15] [Testing]       Re-running... PASSED
[08:26:22] [Done]          T-003 complete. Moving to T-004...
Enter fullscreen mode Exit fullscreen mode

SpecWeave's /sw:auto mode runs autonomously for hours. It implements, tests, fixes failures, and moves to the next task. I've woken up to completed features that would have taken me a full day.

But you need guardrails. Quality gates, test requirements, and clear boundaries on what the agent can and can't change.

4. The documentation writes itself

This is the counterintuitive part. By requiring specs before implementation, you end up with better documentation than any "write docs after" approach. Because the docs ARE the implementation instructions. They're always up to date because the code was generated from them.

Real Usage

I've been dogfooding SpecWeave on my own products:

  • EasyChamp: A sports competition management platform with 14+ microservices. SpecWeave manages the specs, plans, and task breakdown across all services.
  • SkillUp Football: A soccer training app (iOS + Web) with 48 coaches and 235 content pieces. SpecWeave handled the full-stack implementation from spec to deployment.

Both are real, production products serving real users. Not toy demos.

The Numbers

As of this week:

  • 18,500+ monthly npm downloads
  • 43 GitHub stars (still early, but growing)
  • 483 published versions
  • Works with any AI coding assistant (Claude, Cursor, Copilot, Windsurf)

Try It 

npm install -g specweave
Enter fullscreen mode Exit fullscreen mode

Create your first increment:

/sw:increment "Your feature description"
/sw:do
Enter fullscreen mode Exit fullscreen mode

The repo: github.com/anton-abyzov/specweave

Stars appreciated, issues even more so. I'm building this in public and shipping fast.

I'm Anton, an engineer and founder building AI-powered tools and sports tech. I write about building real products with AI on YouTube and LinkedIn. Follow along if you're into spec-driven development or AI coding.

Top comments (0)