Superlearn — плагин

---
name: superlearn
description: Build an interactive learning board on any topic. Use when the user wants to learn, study, or research a subject and get a curated, visual learning experience — researches the live web with the agent's own web search, scrapes YouTube and arXiv for real IDs and papers, runs an iterative research loop into a scratch workspace, authors a validated board JSON, and serves the Superlearn web app locally. Triggers on "/superlearn", "$superlearn", "I want to learn", "teach me", "help me study", "make me a learning board".
---

This is the Superlearn playbook packaged as a standard Agent Skill — the same
pipeline that ships as a Claude Code plugin, usable from OpenAI Codex, Kilo
(formerly Kilo Code), and any agent that reads `SKILL.md`. The playbook below
is written with Claude Code's names for things; resolve these three
differences first, then follow it exactly.

**1. The Superlearn root.** The commands below run bundled scripts as
`"${CLAUDE_PLUGIN_ROOT}/scripts/…"`. Claude Code sets that variable
automatically; elsewhere, define it once in every shell you use:

```bash
export CLAUDE_PLUGIN_ROOT="{{SUPERLEARN_ROOT}}"
```

If the value above is still an unstamped placeholder in curly braces, this
file is inside the Superlearn repo itself: the root is three directories up —
the directory that contains `scripts/validate_board.py` and `app/index.html`.
(`scripts/install.py` stamps the real path when it copies this file into a
global skills directory.)

**2. Tool names.** Map the playbook's tool names to your environment:

| The playbook says | OpenAI Codex | Kilo |
|---|---|---|
| WebSearch | built-in web search (on by default; `--search` for live results) | `websearch` — enable Web Search for non-Kilo providers, or fall back to `webfetch` on known good sources |
| WebFetch | fetch the page from the shell, e.g. `curl -sL <url>` | `webfetch` |
| `superlearn-researcher` agent (Task tool) | spawn sub-agents, one per subtopic, in parallel; a `superlearn-researcher` role ships in `.codex/agents/` | the `task` tool with the `superlearn-researcher` subagent, several concurrently |

If your environment has no subagent support, research the subtopics
sequentially yourself — same steps, same files, one at a time.

**3. Network access.** The scrapers call the YouTube and arXiv APIs and the
app serves on `http://localhost:4321`, so the sandbox needs real network
access. In Codex, that means:

```toml
# ~/.codex/config.toml
[sandbox_workspace_write]
network_access = true
```

In Kilo, allow the `python3` script invocations when prompted (or add a
`permission` rule for them).

Everything else — the phases, the board schema, the quality bar — applies