agent-certifier

# Agent Certifier This skill turns **human certifications** (e.g. PL-300, CPA, SAP B1, Azure AI) into **machine-certifiable agents** with clear skills, benchmarks, and a signed competency contract. Use this skill when the user gives: - A human cert or license name (e.g. "PL-300: Power BI Data Analyst") - Optionally one or more reference repos or products (e.g. `microsoft/powerbi-desktop-samples`) - A target agent name/slug (e.g. `powerbi-bi-architect`) Your job is to emit a **complete, market-ready bundle**: - A skills spec (YAML) with levels, benchmarks, and tools - An agent spec (YAML) wired to those skills - A human-readable `skills.md` for documentation - Optional certification JWT schema, compatible with an external CertificationAuthority --- ## When to Use This Skill Invoke this skill whenever the user wants to: - Translate a human certification (Azure, SAP, CPA, Azure AI, etc.) into an **agent certification ladder** - Define **skills + tools + benchmarks** for an agent in a reusable, model-agnostic way - Produce **ready-to-commit** files for a repo: - `skills/<domain>.yaml` - `agents/<agent_slug>.yaml` - `docs/<agent_slug>-skills.md` The output should be designed so it can be: - Used by Claude Skills (this SKILL.md format) - Loaded by OpenSkills (`anthropics/skills` compatible) - Reused by other agents (Gemini, OpenAI, etc.) via the same YAML contracts --- ## Input Format Assume the user will give you (in natural language, not strict JSON): - Human cert(s) and level(s) - e.g. "PL-300 + DP-500 + Power Platform Solution Architect Expert" - Domain / role - e.g. "Power BI / Fabric BI Architect for retail dashboards" - Reference repos or artifacts (optional but common) - e.g. GitHub URLs, product pages, sample `.pbix`, `.twb`, etc. - Target agent id/slug - e.g. `powerbi_bi_architect` You must **infer missing pieces** safely and document assumptions. --- ## What To Produce Always produce **three main artifacts** (as copy-paste-ready blocks): ### 1. `skills/<domain>.yaml` A YAML file that defines: - `version`, `domain` - `human_analogs`: list of human certs you are mirroring - `sources`: reference repos or artifacts (GitHub, sample files) - `tools`: logical tool contract names (e.g. `pbix_reader`, `dax_analyzer`) - `skills`: - Each with `id`, `level` (fundamentals/role_based/specialty/business/expert etc.) - `human_analog`, `description` - `required_tools`, optional `prerequisites` - `benchmarks`: each with `id`, `description`, `repo_source`, and `success_criteria` list Keep tool names **abstract** so they can be mapped to OpenAI/Gemini/Claude tool schemas later. ### 2. `agents/<agent_slug>.yaml` An agent spec that: - References the domain and skill IDs from the skills YAML - Lists required tools by id - Defines `human_cert_analogs` (strings) - Defines `certification_policy`: - Levels (`fundamentals`, `associate`, `expert`, etc.) - `required_skills` for each level - `min_benchmarks_passed` per level - `issuance` block: - `title_template` - `validity_days` - `conditions` (bullets) Include a `benchmarks_runtime` block describing: - `repo_sources` (e.g. `microsoft/powerbi-desktop-samples`) - `execution.runner` (e.g. `ci.pipeline.powerbi`) - `schedule` (e.g. `nightly`) ### 3. `docs/<agent_slug>-skills.md` A markdown doc for humans that: - Explains which human certs this agent emulates - Lists each skill level with: - Human analog - Capabilities (bullets) - Benchmark(s) and pass criteria - Explains the **certification policy**: - What "Fundamentals / Associate / Expert Certified" means - How the external `CertificationAuthority` JWT is issued & used Structure this as: 1. Overview 2. Human Certification Analogs 3. Tools Required 4. Skill Levels & Benchmarks 5. Certification Policy --- ## Instructions When this skill is active: 1. **Parse the user brief.** - Identify the **domain** (e.g. Power BI, SAP B1, Azure AI, Odoo). - Extract all human certification names and their levels. - Note any reference repos / products / sample files. 2. **Define the skills ladder.** - Map human certs into 3–5 levels: - fundamentals → role_based → specialty → business → expert - For each level: - Write a concise description of capabilities. - Choose the tools needed (abstract names). - Design 1–3 concrete benchmarks that can be evaluated automatically. - Benchmarks must have **clear, measurable pass criteria** (e.g. KPI parity within 1%, build succeeds, tests green, etc.). 3. **Design tool contracts.** - Keep tools **model-agnostic**: - Do NOT hard-code OpenAI / Gemini / Claude APIs. - Use logical names (e.g. `ocr_gateway`, `sap_b1_api_client`, `pbix_reader`). - Focus on what the tool does, not how it's implemented. 4. **Generate the three artifacts.** - Emit them in this order, each in its own fenced code block: 1. `skills/<domain>.yaml` 2. `agents/<agent_slug>.yaml` 3. `docs/<agent_slug>-skills.md` - Make them **ready to commit** (no placeholders like `TODO` or `...`). 5. **State assumptions & risks briefly.** - At the end, add a short "Assumptions & Risks" section (3–6 bullets) outside the code blocks: - List any big assumptions (e.g. data source, repos, tools). - Flag anything that absolutely needs human review (compliance, PII, prod access). --- ## Examples ### Example 1 – Power BI Architect > "Create an agent that is equivalent to PL-300 + DP-500 + Power Platform Solution Architect, using `microsoft/powerbi-desktop-samples` as the benchmark repo. Agent slug: `powerbi_bi_architect`." You should: - Define `domain: "cloud_ai_platforms.power_bi"` - Map the certs into a skills ladder (fundamentals → expert) - Use Store Sales / Competitive Marketing Analysis `.pbix` as benchmarks - Emit YAML + docs as specified above ### Example 2 – Azure AI Engineer > "Build an agent certified at the same level as Azure AI Engineer Associate, focused on RAG systems for finance dashboards." You should: - Create a `cloud_ai_platforms.azure_ai_rag` skills domain - Define tools like `embedding_indexer`, `rag_query_runner`, `azure_openai_client` - Add benchmarks: end-to-end RAG flow, latency, accuracy, hallucination checks - Emit all three artifacts. ### Example 3 – Retail Analytics (Scout) > "Certify an agent at the same level as a Retail Analytics Data Engineer for sari-sari / FMCG dashboards, using the Scout dashboard as the benchmark." You should: - Create `retail_analytics.scout` domain - Define tools: `supabase_query_runner`, `kpi_validator`, `chart_renderer` - Add benchmarks: schema validation, KPI accuracy, dashboard render time - Emit YAML + docs --- ## Guidelines - Prefer **clear, testable benchmarks** over vague descriptions. - Keep everything **implementation-agnostic**: - No hard-wiring to a single model provider. - Assume tools can be backed by Claude, OpenAI, Gemini, or local models. - Favor **production-readiness**: - Think like a vendor shipping a marketplace agent, not a demo. - Include governance/security considerations where relevant (RLS, PII, secrets). - Never include real secrets or API keys in outputs. --- ## Related Skills - `engine-spec-writer` – Creates engine.yaml specs for IPAI Platform - `supabase-schema-designer` – Designs Supabase schemas with RLS - `dashboard-builder` – Builds Next.js dashboards from specs