schematize
Skills · Claude Code · Agent Skills

AI engineering the agent follows — no negotiation.

Build with LLM on the house's same security and data floor: prompt injection is a first-line attack, output untrusted until validated, no LLM decides authorization and RAG isolated per tenant.

v0.1.0IA · LLM/RAG/Agentesgithub.com/schematizeme/skill-ai

تتم صيانته بواسطة Lucas Andrade

schematize-ai · v0.1.0bash
$ schematize install ai
GitHub →تنزيل المهارة (.zip)/ai-help · github.com/schematizeme/schematize-cli

Non-negotiable floors

Never switched off "to go faster". Full list in the skill.

No LLM decides authorization

The model suggests; authorization is deterministic server-side code (house IAM, deny-by-default). tenant_id/role come from the token, never from the prompt.

AI output untrusted until validated

It only enters the system after deterministic validation (schema, allowlist, range). Never eval/SQL/shell/raw HTML straight from the LLM output.

Prompt injection is a first-line attack

Every text in the prompt — user, RAG, tool, web — is hostile until proven otherwise. An instruction coming from data is not an order.

Secrets never in the prompt or the client

Provider key and credentials stay out of the prompt, the context and the browser. The LLM call is server-side; the client talks to your backend.

Provider-agnostic, RAG per tenant

Model access behind a pluggable interface (Claude default); switching is config + ADR. Vector index partitioned per tenant, with source citation.

No eval, not done

Golden dataset, regression and injection/jailbreak red-team in CI. Trace per call (tokens, cost, latency) and budget from day 1.

Commands

Installed in .claude/commands/. Run /ai-help anytime.

/ai-helpLists every command.
/ai-loadForce-loads the whole normative body: LLM engineering, RAG, agents, evals/guardrails, security, observability.
/ai-ragDesigns/audits a RAG pipeline: ingestion → chunking → embeddings → retrieval → evaluation, with per-tenant isolation and source citation.
/ai-evalBuilds/runs the evals suite: golden dataset, calibrated LLM-as-judge, regression, quality gate.
/ai-guardrailsAudits/scaffolds the guardrails: schema output validation, deny-by-default, prompt injection/jailbreak red-team, PII/safety.
/ai-claudeCreates or merges the always-on AI-engineering CLAUDE.md at the repo root.
/ai-ccContext compact: writes the handoff to the archive and runs /compact.
/ai-handoffWrites the handoff (context.md + checklist.md) without compacting.

For AI agents

Install with no human in the loop. Metadata in llms.txt.

  • Run: git clone https://github.com/schematizeme/skill-ai.git && ./skill-ai/install.sh
  • Or download: https://github.com/schematizeme/skill-ai/releases/latest/download/skill-ai.zip and unzip into .claude/skills/
  • Copy assets/CLAUDE.md to the repo root (always-on rule).
  • Confirm with /ai-help.

لا تريد التثبيت؟

اقرأ المعايير الكاملة وانسخها مباشرة إلى مشروعك — أو ثبّت المهارة ليفرضها الوكيل نيابةً عنك.

GitHub →الحزم