apastra（apastra 命令行）

Apastra PromptOps Skills

Skills for managing AI prompts with the same discipline as code — versioned files, automated evaluations, regression detection, and baseline tracking. Your IDE agent is the harness.

Installation

This is a collection repository containing multiple skills. Install individual skills by path:

# Start here — onboarding and project setup
npx skills add BintzGavin/apastra/skills/getting-started

# Core workflow skills
npx skills add BintzGavin/apastra/skills/eval
npx skills add BintzGavin/apastra/skills/baseline
npx skills add BintzGavin/apastra/skills/scaffold
npx skills add BintzGavin/apastra/skills/validate

# Or install everything at once
npx skills add BintzGavin/apastra --all

Available Skills

Getting Started

• skills/getting-started — Project setup, first prompt spec, first evaluation. Start here.

Workflow Skills

• skills/eval — Run evaluations using your IDE agent as the harness. Compare results against baselines.
• skills/baseline — Establish and manage evaluation baselines for regression detection.
• skills/scaffold — Generate new prompt specs, datasets, evaluators, and suites from templates.
• skills/validate — Validate all promptops files against JSON schemas.
• skills/setup-ci — Upgrade from local-first evaluation to automated GitHub Actions CI.

When to Use

Use these skills when:

• Managing AI prompts that need version control and quality gates
• Running evaluations locally without CI or cloud infrastructure
• Detecting prompt regressions before shipping
• Establishing baselines for prompt quality tracking
• Scaffolding new prompt specs, test datasets, or evaluator definitions

Repository

View all skills and source code at: https://github.com/BintzGavin/apastra

🤖 OVERARCHING CONTEXT FOR AI AGENTS You are operating within Apastra, a state-of-the-art PromptOps architecture. The Core Philosophy: Prompts must be treated like disciplined software assets. Apastra is a file-based protocol. There is no hidden database, no mandatory SaaS platform, and no forced framework. - Durable state lives in Git: Prompts, test datasets, evaluator scoring rules, benchmark suites, baselines, and regression policies are all defined as flat files (YAML/JSON). - The Agent is the Harness: You (the external IDE agent) perform the computation. You read the suites, run the test cases, apply the scoring rules, and compare the results to the baselines. - Append-only Artifacts: When you run evaluations, you generate run manifests, scorecards, and evidence records. You never overwrite a baseline; you archive and append. Local-First vs. CI Upgrades: Developers using this project prefer a zero-infrastructure local-first approach. They evaluate prompts locally using you as their test runner. However, if they need enterprise-grade governance, they can upgrade to automated pull request gating using the apastra-setup-ci skill. The file formats remain exactly the same. As an external agent using these skills, your job is to enforce this discipline. Never mutate a file ad-hoc if there is a structured prompt spec for it. If you are told to "evaluate a prompt", follow the exact file-based pipeline defined in the apastra-eval skill.