CurdX Flow

Spec-driven delivery for Claude Code. Turn "build this feature" into reviewable specs, traceable tasks, and proof — instead of "it says it's done."

npm exec -- @curdx/flow@latest install curdx-flow --yes

/curdx-flow:start todo-app  build a todo app with create/edit/delete, browser-verified

That's it. The plugin classifies the goal, writes specs into your repo, executes task-by-task, and refuses to call it done without evidence.

The problem it kills

Three failure modes show up the moment Claude Code works on something real:

Without Flow	With Flow
Context rots across turns — the model forgets the original ask.	Goal pinned to requirements.md. Survives every restart.
"Done" arrives without proof — no command output, no browser run.	Completion requires verificationBlocks. No silent passes.
Big features either crush small flows or are tackled in one chaotic pass.	/start auto-routes: quick fix, light spec, full spec, resume, or triage into multiple specs.

This isn't another planning ceremony. It's an execution discipline layer for Claude Code.

What lands in your repo

Every real feature run ends with these artifacts — committable, reviewable, resumable:

specs/todo-app/
├── research.md         # facts, risks, existing assumptions
├── requirements.md     # what to build · what counts as done
├── design.md           # how to build · what's in scope
├── tasks.md            # bounded steps, each with a verify command
├── .curdx-state.json   # phase, verificationBlocks, recovery
└── .progress.md        # runtime breadcrumbs (usually gitignored)

Plus the code diff. Plus test/build/browser/CI/release evidence stored in state. The PR carries the goal, the plan, and the proof — not just the diff.

3-minute tour

# 1. Install (also installs companion plugins + ~/.claude/CLAUDE.md block)
npm exec -- @curdx/flow@latest install curdx-flow --yes

# 2. Open your project in Claude Code
cd /path/to/your/project && claude

Inside Claude Code:

/curdx-flow:start todo-app build a todo app with create/edit/delete, browser-verified
/curdx-flow:status       # → "ready to /implement"
/curdx-flow:implement    # runs tasks, captures evidence

Lost mid-session? Run /curdx-flow:status — it tells you the active spec, phase, missing capabilities, and the exact next command.

Why teams adopt it

• Reviewable AI work. requirements.md + design.md go into the PR. Reviewers see intent, not just diffs.
• Resumable across sessions. Close the laptop, come back tomorrow — status and start reattach to the spec.
• Provable completion. Tests, build, browser DOM, CI runs, npm releases — all captured into verificationBlocks.
• Right-sized process. Tiny edit? Direct execution. Cross-module epic? Triage into linked specs.
• No vendor lock-in. It's a Claude Code plugin. Specs are plain Markdown. Walk away anytime.

Best fit · not a fit

Use it when	Skip it when
Frontend page, CLI, plugin, API, or release work	One-shot Q&A — "what does this snippet do?"
The task needs requirements before code	Throwaway script with no review path
You want browser/test/CI/release evidence	You explicitly want answers without file edits
The work may pause and resume mid-session	A 1-line fix where ceremony is dead weight
You want PRs that carry plan + proof	—

Pitch it in one breath

CurdX Flow turns a Claude Code prompt into spec files, traceable tasks, and verified evidence — so AI-assisted features become reviewable, resumable, and trustworthy, not just "it says done."

npm exec -- @curdx/flow@latest install curdx-flow --yes

Next

• Getting Started — your first run, end-to-end.
• Commands — what to type, when.
• How It Works — the mental model and the internals.
• Configuration — flags worth knowing.
• Troubleshooting — when something looks off.