The workflow assumes you scaffolded with sn init. The scaffolded CLAUDE.md is the AI-collaboration contract — it tells Claude which operations exist, that every step must use upsert-style idempotency, and that Claude should check schemas before writing steps. You write the feature description; Claude does the translation.
Setting Up Claude
Point Claude Code at your scaffolded project
The typical loop: describe your goal to Claude → Claude generates a manifest → you review it → run sn execute → sn validate → sn test. Claude handles the mechanical work; you stay in control of what gets deployed.
You can extend the scaffolded CLAUDE.md with your instance's naming conventions, app-scope naming rules, or team-specific rules. Don't replace it — the baseline rules are what keep generated manifests passing preflight.
Tasks
1
Install Claude Code if you haven't already.
curl -fsSL https://claude.ai/install.sh | bash2
Open Claude Code in the project root. Claude reads the scaffolded `CLAUDE.md` on startup.
cd my-sn-project && claude3
Describe a feature in plain English. Claude follows the rules in `CLAUDE.md` to produce a manifest.
Create a business rule on the incident table that sets priority to 4 when no priority is specifiedI'll create a manifest for that. Let me check the schema first...
$ sn ops business-rule create
Here's the manifest:
✓ Created manifests/default-priority/manifest.ts
- 1 step: business-rule > create
- Table: incident
- Trigger: before insert
Run `sn execute manifests/default-priority/manifest.ts` to deploy it.