Free skills. Inspectable source. Your own workspace.
A practical toolkit for learning with your agent: shape a workflow, question its evidence and improve what you build. Install once, then choose the work that matters to you.
Prefer to begin without setup? Browse the free learning paths →
Package name: starlight-department-lab. Your agent tool may have its own model or subscription costs.
Practice in your conversation
Connect the Academy's public teaching methods and challenges to Codex or Claude Code. ChatGPT setup and its interactive card are also available to inspect, with host verification still pending.
Explore the Academy connection →Bring your own agent
Reviewed public source
Add the reviewed marketplace revision, then install its Academy plugin. Your host handles trust and permissions.
You need Git, Node.js 22 or later, and a current Codex CLI. Review the source, then run both commands in order.
codex plugin marketplace add frankxai/starlight-academy-plugin --ref 60724543685f6a5db7cbed55fe4b4aa574913ff2
codex plugin add starlight-department-lab@starlight-academyAlready using a local Starlight marketplace? Inspect its source in your host's marketplace manager before replacing it. Keep practice files outside the plugin cache. A pinned revision stays fixed until you explicitly select an update.
Manifest validated; root-resolved MCP handshake checked. Installation and full host startup are separate checks.
Clean Git installation and the installed MCP handshake were checked on 2026-09-12T12:01:43Z. Inspect the exact public revision ↗
Read Codex's official plugin documentation ↗codex plugin remove starlight-department-lab@starlight-academyYour saved practice files remain separate from the plugin.
Your first useful artifact
Agentic Systems
Turn one recurring workflow into a small, inspectable operating contract.
Expected artifactAGENT_SYSTEM_CONSTITUTION.mdKeep the first attempt, critique and revision visible. Then test your method in a changed context.
Use the learn-starlight-mission skill. Load get_mission_studio with missionId "systems.bound-one-agentic-workflow".
Guide me through "Bound one agentic workflow" and help me produce AGENT_SYSTEM_CONSTITUTION.md.
First ask for my own attempt before offering hints or reference answers. Use public, synthetic or redacted examples. Preserve my assistance declaration, sources, artifact, critique, revision and transfer attempt. Explain the limits of each check.
If I bring an existing practice file, use inspect_mission_practice to check its compatibility before continuing. Preserve its original first attempt, assistance declarations and critique snapshots; do not restart or overwrite them. If it is incompatible, explain the mismatch and keep the original file.
Export files to my workspace only when I ask. This is self-study and grants no credential or operational authority.If your agent cannot find get_mission_studio, check that Node is visible to the app, restart the session, and inspect its MCP logs.
Copying a command does not install the plugin. Your agent tool confirms installation.
Download and extract the complete learning pack. With Node.js 22 or later and your chosen tool installed, open a terminal in the extracted starlight-academy folder and run its command:
node install.mjs codexnode install.mjs claudenode install.mjs groknode install.mjs hermesnode install.mjs antigravityThe included README explains native trust prompts and updates. These instructions remain readable when JavaScript cannot run. Read the mission contracts in Mission Studio to choose your first exercise.
Inside the toolkit
A skill is a readable procedure your agent can follow. These five ship together. Browse what each does, try related browser practice, or inspect its instructions before installing.
Ask your agent to coach you through a mission while keeping your own first response intact.
You work toward A portable practice record with sources, drafts and review history.
After installing, ask it to use learn-starlight-mission. Give it a public or synthetic example, and ask what inputs it needs before proceeding.
A skill can guide an agent; it does not prove the agent followed the procedure. Keep the evidence and seek independent review when the work matters.
Give one recurring workflow a purpose, named roles, a review moment and a human owner.
You work toward A department packet you can inspect and revise.
After installing, ask it to use design-ai-department. Give it a public or synthetic example, and ask what inputs it needs before proceeding.
A skill can guide an agent; it does not prove the agent followed the procedure. Keep the evidence and seek independent review when the work matters.
Check a draft against its declared structure and review requirements before trusting it.
You work toward Explicit checks and issues to resolve; no deployment authority.
After installing, ask it to use verify-ai-department-package. Give it a public or synthetic example, and ask what inputs it needs before proceeding.
A skill can guide an agent; it does not prove the agent followed the procedure. Keep the evidence and seek independent review when the work matters.
Judge synthetic agent cases before revealing their reference choices. Explain what the evidence supports.
You work toward Your first judgment, a comparison and a proposed follow-up test.
After installing, ask it to use practice-agent-evaluation. Give it a public or synthetic example, and ask what inputs it needs before proceeding.
A skill can guide an agent; it does not prove the agent followed the procedure. Keep the evidence and seek independent review when the work matters.
Explore acceptance fixtures and distinguish missing evidence from a passing result.
You work toward A fixture result with visible limits, ready for your own review.
After installing, ask it to use test-workflow-evidence. Give it a public or synthetic example, and ask what inputs it needs before proceeding.
A skill can guide an agent; it does not prove the agent followed the procedure. Keep the evidence and seek independent review when the work matters.
If a full mission feels like too much, begin with one judgment:
Use the practice-agent-evaluation skill. Show me the confident-completion case without its reference answers. Preserve my first judgment, then compare my choices, explain the evidence and ask me to design a timeout test. Keep the result labelled as self-study.
Open a fresh agent session after installing. The MCP should expose eight tools, including get_evaluation_cases, check_evaluation_answers and run_acceptance_suite. The checker compares public reference choices; it does not grade written reasoning or a live agent run.
The local server reads bundled public and synthetic material. It makes no network or model calls and writes no files. Your host agent's tools and permissions are separate.
The archive contains readable source, native manifests, the installer and per-file checksums. SHA-256 checks consistency; it is not a publisher signature. Extract the complete archive and keep the directory for local marketplace updates.
e857a17af797cd67c9cd052d03667106503d2341c8b0cb177807cce0eab07e7fA website download cannot silently install into your desktop agent. Each host controls installation and trust. Use the verified Git commands where available, or the downloadable pack's host-specific adapter. Installing this package does not imply admission to a provider's public directory.
No hosted tutor, paid model plan, credential or operational authority is included. The toolkit supports self-study; its public answer keys are not a secure examination.