blueprint-figure
Draws static SVG figures for lessons, READMEs and manuals in the AI Engineering from Scratch look: paper panel, blueprint accent,…
Build durable understanding by finding a learner's frontier, agreeing on a dependency route, developing one connection at a time, and checking each connection. Use for tutoring, studying, guided practice, and explanations meant to be retained.
Claude Code
$ npx -y skills add Zproger/ConceptLoom --skill concept-coach --agent claude-code
Codex
$ npx skills add Zproger/ConceptLoom --skill concept-coach -a codex -y
Cursor
$ npx skills add Zproger/ConceptLoom --skill concept-coach -a cursor -y
Gemini CLI
$ npx skills add Zproger/ConceptLoom --skill concept-coach -a gemini-cli -y
GitHub Copilot
$ npx skills add Zproger/ConceptLoom --skill concept-coach -a github-copilot -y
OpenCode
$ npx skills add Zproger/ConceptLoom --skill concept-coach -a opencode -y
Windsurf
$ npx skills add Zproger/ConceptLoom --skill concept-coach -a windsurf -y
name: concept-coach description: Build durable understanding by finding a learner's frontier, agreeing on a dependency route, developing one connection at a time, and checking each connection. Use for tutoring, studying, guided practice, and explanations meant to be retained.
Teach so the learner can recreate the conclusion from ideas they trust.
Read [the coaching reference](references/coaching-reference.md) before beginning a learning session. Follow Locate, Weave, and Build without collapsing them into a generic explanation.
Learning sessions must survive restarts. Start a notebook and persist progress with `loom_save_learning_state` throughout the lesson. If the learner asks to continue or resume, call `loom_load_learning_state` before responding with lesson content.
After every learner response, persistence must be your next action. Save route approval as `phase=build` before asking the first teaching question. Assess checkpoints with `sessionId` and an exact `nextStep`; assessment saves progress in the same call. Never send the next lesson message first.
Use `AskUserQuestion` for preferences and for presenting checkpoint choices. Grade only knowledge questions, using `loom_frame_checkpoint` before asking and `loom_assess_checkpoint` after the learner selects keys.
Choose evidence-appropriate checkpoints: choice for recognition, free recall for reconstruction, prediction or debugging for application, and transfer for a new context. Record confidence when useful and pass a stable `evidenceConcept`. At resume, check `loom_list_due_reviews` before reteaching saved material.
Delegate source verification to the `evidence-scout` subagent when facts are uncertain or time-sensitive. Prefer a compact text diagram; use a Mermaid block only when it makes a larger relationship clearer. Wait for learner approval after proposing the dependency route.
A portable AI learning workflow for Codex, Claude Code, OpenCode, and DeepSeek Harness. Concept Loom turns a coding agent into a structured learning partner.
Draws static SVG figures for lessons, READMEs and manuals in the AI Engineering from Scratch look: paper panel, blueprint accent,…
End-to-end workflow for creating complete JavaScript concept documentation, orchestrating all skills from research to final review
Test a learner on a single Claude Code tutorial lesson (01-10) with 10 questions, scoring answers and flagging weak spots. Use…