Keep AI coding decisions in git
AI coding agents (GitHub Copilot, Claude Code, Grok, Ollama, and similar tools) can change a lot of code before lunch. By the next day, the hard parts—why that edge case was “fine,” what was out of scope, what “done” meant—are often only in a chat thread.
That is not a model problem. It is a process and storage problem.
Local Spec-Driven Development means: short specs and progress live in your repository, next to the code, so humans and agents read the same files. sdd is a CLI that runs that process without a cloud process product or a required IDE extension.
The failure mode of chat-only coding
| What happens | Why it hurts |
|---|---|
| Decisions stay in the agent chat | New chats (or new teammates) start from zero |
| PRs show diffs, not intent | Reviewers re-discover the “why” from code alone |
| Context windows fill with source | Plans get compressed away; constraints are forgotten |
| Process is “vibe” until it isn’t | Hotfixes and platform work need different depth |
If this feels familiar, you do not need a heavier enterprise ALM suite for every three-line fix. You need a light spine in git.
What local Spec-Driven Development looks like
- Start a piece of work (
sdd new "…", orsdd greenfieldfor a new product). - Fill a short markdown file (intent, design, tasks—not a novel).
- Advance with
sdd nextwhen the step is real. - Implement, verify, then
sdd complete.
Notes for this PR live under changes/. Product truths that should last live under memory/ (constitution, backlog after greenfield, thin architecture).
Chat / agent window Your repo (source of truth)
│ │
│ sdd handoff + packs │
└─────────────────────────────►│ changes/ · memory/ · .sdd/Agents get a handoff and host rules; you still own the process with the CLI.
How sdd fits (and what it is not)
| sdd is | sdd is not |
|---|---|
A Spec-Driven Development CLI (npm i -g @structured-vibe-coding/cli) | A cloud ticket system |
| Process coach for you + your AI coding agent | A replacement for your framework or IDE |
| Markdown in git, progressive workflows (hotfix → enterprise) | One rigid ceremony for every change |
| Optional MCP sources sdd calls (design system, AST, APIs) | An inbound “drive sdd next from Cursor tools” server (by design) |
Who this is for
- Solo and small teams using agents daily
- Teams that want decisions reviewable in PRs
- Greenfield MVPs that need a vision → backlog → F-001 path
- Groups that sometimes need gates without living in them
Not for: people who want process only in a SaaS UI, or who refuse any markdown next to code.
Try it in about 10 minutes
npm install -g @structured-vibe-coding/cli
cd your-app
sdd init --here --ai copilot # or grok | claude | ollama
sdd new "Fix empty list crash" -w hotfix -y --no-agent
# write a few real sentences in the intent file
sdd next --no-agent
sdd complete --no-agentStep-by-step: Start in 10 minutes.
Product concept: What is sdd?.
Why this exists: Why sdd exists.
Related
- Change packs & memory — what to keep forever vs per PR
- Set up your AI — Copilot, Claude Code, Grok, Ollama
- Greenfield — new product from one sentence
- CLI reference — commands