SpecForge

💡 Inspiration

Every developer knows the pain: someone creates a GitLab issue that says "Fix the login flow" — and that's it. No acceptance criteria. No edge cases. No complexity estimate. Teams then waste hours in refinement meetings filling in details that AI can handle in seconds.

I built SpecForge to eliminate that waste entirely and give every developer an instant spec-writing teammate right inside GitLab.

🔮 What it does

SpecForge transforms vague GitLab issues into sprint-ready specifications in seconds — directly inside GitLab Duo Chat.

Trigger: A developer opens any GitLab issue (however vague) and invokes SpecForge in the GitLab Duo Chat sidebar.

Agent response: SpecForge instantly generates a full specification including:

• ✅ User stories — who needs this and why
• ✅ Acceptance criteria — clear, testable conditions
• ✅ Edge cases — what could go wrong
• ✅ Definition of done — what "finished" actually means
• ✅ Complexity estimate — XS / S / M / L / XL sizing

No more refinement meetings. No more guessing. Just open an issue, chat with SpecForge, and get a professional spec in seconds.

🛠️ How I built it

• GitLab Duo Agent Platform — agent orchestration and GitLab Duo Chat integration
• Anthropic Claude (via GitLab) — intelligent spec generation and reasoning
• GitLab Flows — workflow automation and trigger handling

The agent is invoked directly from GitLab Duo Chat inside any issue. It reads the issue context, reasons about what a complete specification should include, and returns a structured, actionable spec — all without leaving GitLab.

🚧 Challenges I faced

The biggest challenge was prompt engineering: getting SpecForge to produce specifications that are genuinely useful — not generic, not too verbose — but crisp, developer-ready output every time. It required iterating heavily on how Claude interprets vague issue text and structures the response.

Integrating cleanly with the GitLab Duo Agent Platform and ensuring the agent responded naturally inside the chat sidebar was also a key technical challenge.

📚 What I learned

• How to build and deploy real agents on the GitLab Duo Agent Platform
• The power of pairing GitLab Flows with Claude for contextual AI responses
• That the best AI tools are invisible — they fit into existing workflows (GitLab issues) rather than forcing new ones

🚀 What's next

• Auto-trigger SpecForge when a new issue is created (zero manual steps)
• Push the generated spec back directly as an issue description update
• Add team-specific templates so specs match your squad's exact format
• Complexity estimate integration with sprint planning tools

Built With

• agent
• anthropic
• claude
• duo
• flows
• gitlab
• platform
• yaml