SGAI Goal Authoring
Compose GOAL.md files for SGAI workspaces through interactive conversation. This skill detects your tech stack, selects the right developer and reviewer agents, and interviews you to build a complete goal file with an agent flow graph, model preferences, a completion gate, and outcome-based success criteria.
Installation
Add the marketplace and install this skill:
/plugin marketplace add britt/claude-code-skills
/plugin install sgai-goal@britt
Or install all skills at once:
/plugin install claude-code-skills@britt
Compatibility
| Platform | Supported |
|---|---|
| Claude Code | ✓ |
| Claude.ai |
How to use it
Use this skill when:
- Starting a new feature, project, or bug fix in a repository that uses SGAI
- Asked to “create a goal” or “set up SGAI”
- Describing work to be done in an SGAI-managed repo
The skill first checks for an .sgai/ directory (or other SGAI configuration) and stops to confirm with you if it can’t find one.
Features
Automatic tech stack detection
Scans the repository for go.mod, package.json, wrangler.toml, vercel.json, and other indicators to choose the right developer and reviewer agent pairing — Go, React, HTMX, shell/CLI, Cloudflare Workers, Vercel, and Agent SDK stacks are all recognized, and multiple stacks combine into a single flow.
Model preference memory
Reads saved model preferences from .sgai/preferences.json, or asks once and saves the answer for future goals.
Outcome-based success criteria Turns a plain description of the work into checklist-style success criteria that describe outcomes (“users can sign in with OAuth providers”) rather than implementation steps.
Agent flow graph composition
Builds the developer -> reviewer -> stpa-analyst flow graph for the detected stack, always terminating in stpa-analyst for safety analysis except in pure documentation flows.
Completion gate detection
Infers a completionGateScript from the project’s test tooling (go test ./..., npm test, make test) so SGAI knows how to verify the goal is done.
Safe overwrite handling
If a GOAL.md already exists, it is renamed to a timestamped backup before the new one is written — no work is lost.