Specwright
Live · v1.7.0Spec-driven workflows for Claude Code. 14 commands, 6 subagents and 7 guard-rail hooks that stop the model from coding before you approve.
MIT · Windows, macOS, Linux
View on GitHub// spec-first ai delivery
AI writes the code. Specs keep it honest.
DevelopZone builds open-source tooling that puts specs, gates and reviews around AI coding assistants, and helps software teams adopt them.
Open source · MIT · Built on Claude
claude · specwrightgate 1 of 3
> /sd:feature todo-priority
Phase 1 sd-spec-architect
-> .specs/FEAT-todo-priority/00-spec.md
Success criteria AC-1 .. AC-9
Out of scope 6 items
Constitution layer rules OK · errors OK
Gate 1 · Spec approval [STOP]
Approve spec? (yes / refine / abort)
_
// the problem
AI coding is fast. Unreviewed AI coding is expensive.
Without gates
- Diffs arrive without the intent behind them.
- Invariants are forgotten between sessions.
- Fixes land on the symptom, not the cause.
With a spec-first process
- Every change starts as a written spec you approve.
- Gates halt the workflow until a person answers.
- Reviews come from an agent that cannot write code.
// products
Three tools, one discipline.
Graphwright
Coming soonA Roslyn-native code graph for .NET, served over MCP: symbol search, references and call graphs your AI assistant can query.
Open source · .NET · MCP server
.NET Migration Assessor
Coming soonPoint it at a .NET Framework or .NET 8 solution. Get a ranked, explained migration plan to .NET 10.
Powered by Graphwright and Claude
Join the waitlist// how specwright works
Gate 1 is a stop, not a suggestion.
- 01Gates halt the workflow. Silence is not approval. A phase does not advance without an explicit answer.
- 02The block lives outside the prompt. A PreToolUse hook denies Edit and Write until a spec is approved. A prompt can be argued with; a hook cannot.
- 03The reviewer cannot write. It has no write tools, so every finding goes back through a fresh implementer.
- 04Specs are inputs, not write-ups. Every subagent is handed the spec on invocation, so it cannot rot into documentation nobody reads.
14
slash commands
6
specialized subagents
7
guard-rail hooks
6
spec templates
Stack-agnostic: .NET, Node, Python, Go, Rust. Runs on Windows, macOS and Linux.
// built on claude
Claude does the reasoning. The process does the governing.
today
Specwright on Claude Code
Sonnet writes specs, plans and reviews. Haiku executes one atomic task at a time. Model aliases, never pinned IDs, so workflows move forward with each release.
next
Migration Assessor on the Claude API
Roslyn extracts the facts. Claude explains each breaking change and estimates the effort, with prompt caching and batch runs for large solutions.
your .slnGraphwright · factsClaude · reasoningranked migration plan
// services
Bring AI into your delivery process, with guard rails.
01 · workshop
Spec-driven AI coding, hands on
A half-day session for your team: writing specs, setting gates and reviewing AI output on your own codebase.
02 · setup
Claude Code in your repository
Claude Code and Specwright installed, with a constitution written for your architecture and conventions.
03 · review
An audit of how you use AI today
Where your team's AI workflow is fast, where it is risky, and which gates would pay for themselves.
// founder
Trong Tran
12+ years building .NET backends. I started DevelopZone to make AI-assisted development auditable instead of magical.
DevelopZone is an early-stage, one-person company based in Vietnam. Everything on this page is real and current, including what is not shipped yet.