sonzai-claude-skill
A multi-plugin skill repo for AI coding agents (Claude Code, Codex, Gemini CLI, Copilot CLI) building on the Sonzai SDK. Two plugins ship from this repo:
| Plugin | Skills | Audience |
|---|
sonzai-sdk (public) | sonzai-sdk (wizard) + full-auto (autonomous) + cto-loop (semi-auto with 2 operator gates) | Any developer using the Sonzai SDK |
sonzai-internal-staff (internal-only) | sonzai-internal-staff | Sonzai staff with read access to private monolith repos |
The sonzai-internal-staff plugin is install-time gated — it lives in a separate plugin manifest, so it is never copied to a non-staff disk just because they installed sonzai-sdk.
Which skill should I use?
| If you have... | And you want... | Use |
|---|
| A scoping conversation with stakeholders | Interactive wizard → spec + plan (no build) | sonzai-sdk |
| A meeting transcript + no human in the loop | Unattended build → running app, no prompts | full-auto |
| A meeting transcript + a tech lead available | Build + 2 operator gates (masterplan + live-app review) | cto-loop |
full-auto and cto-loop share the same core machinery (tech-stack derivation/intake, brownfield audit, version-search hard rule, builder + fixer subagents, docker-compose deploy, functional QA loop). The only differences:
full-auto runs end-to-end autonomously; auto-fixer loop on QA failure (bounded 5 cycles).cto-loop inserts two gates (masterplan approval + live-app review), swaps autonomous tech-stack derivation for an interactive 7-question intake on greenfield (or a detect-and-confirm audit on brownfield), and replaces the auto-fixer with an operator-feedback-driven fixer (bounded 5 cycles + 1 explicit "one more").
What the skills do
sonzai-sdk skill (the wizard)
When invoked:
- Step 0 — drift check. Compares the installed SDK version's committed OpenAPI snapshot against the live spec at
https://api.sonz.ai/docs/openapi.json. Catches stale code generation.
- Step 1 — runs the wizard (
intake.md). 8-question diagnostic interview (Q1–Q8: archetype, integration path, latency, capabilities, brand/persona, proactive, scope, runtime mode). Skips any answered by workspace inference. Forks on greenfield vs existing-codebase.
- Loads the archetype playbook that matches the developer's intent — one of:
companion — 1:1 persistent companion (Replika-shaped)guide-router — MBTI / personality-routed (intake guide + N specialists)enterprise-assistant — team-shared, KB-heavy, auditedcustomer-support — KB + custom tools + webhooksgame-npc — inventory + custom states + dialogue + eventscoach-therapist — long sessions, sync memory, diary, moodhybrid-custom — combinations / multi-tenant
- Drives spec → review → plan via the playbook's templates. Output is a fully-specified implementation plan ready to execute via
superpowers:subagent-driven-development or superpowers:executing-plans.
For developers who just want a syntax lookup, a skip-wizard path falls through to the per-language references (python.md, typescript.md, go.md).
full-auto skill (autonomous, transcript → running app)
Take a transcript, ship a running local app. Zero operator prompts. Bounded fix loop.
- Build-complete notifications (v1.7.0+): sends a Slack DM + Gmail when the build finishes (healthy or with failing QA). Useful for long unattended runs.
- Pre-flight — drift check (live OpenAPI), Docker check,
SONZAI_API_KEY check
- Phase 0a-d — transcript analysis → project-type detection → tech-stack derivation (greenfield) OR brownfield audit → derive 8 sonzai wizard answers
- Phase 1 — assemble masterplan at
docs/cto-review/<date>-masterplan.md
- Phase 2 — builder subagent (always-search-current-state hard rule; never hallucinates package versions)
- Phase 3 — generate Dockerfile +
docker-compose.yml (postgres / postgres+redis / stateless variants); docker compose up -d --wait; functional QA against the running app
- Phase 4 — auto-fixer loop on QA failure (bounded 5 cycles)
- Phase 5 — final report at
docs/cto-review/<date>-final-report.md; live app stays running
Operator owns the push decision — full-auto commits locally, never pushes.
cto-loop skill (semi-auto, transcript → running app + 2 gates)
Same pipeline as full-auto, with two operator gates inserted and an interactive tech-stack intake on greenfield (instead of autonomous defaults).
- Async replies via Slack/Gmail (v1.7.0+): when Gmail + Slack MCPs are enabled, operator can approve/feedback/abort gates from their inbox or Slack DM instead of returning to the terminal. 20min poll cadence, 24h timeout, paused-resumable.
