Lab Homework

Lab Homework

Generates a personal practice assignment tied to a specific Claude Code Lab meeting.

Language

Default: English. Switch to Russian if the user wrote to you in Russian. The generated assignment file is written in the same language as the user's request.

Inputs

1. Curriculum manifest

Fetched from the public API:

https://agency-lab.glebkalinin.com/api/curriculum/<cohort>.json

• Cohort slug stored in <vault>/.config.md under cohort:. On first run, ask the user via AskUserQuestion — options: goal-driven-design-ru, goal-driven-design-en, claude-code-lab-05, claude-code-lab-04 (older cohorts claude-code-lab-03/-02 on request).
• Cache: write the response to <vault>/.cache/curriculum-<cohort>.json with an added top-level _fetched_at ISO timestamp. Reuse the cache if younger than 24h.
• Refresh triggers: user phrases "refresh curriculum", "update curriculum", "pull latest", "обнови программу", "обнови курс" force a re-fetch. If the cache is older than 7 days, offer a refresh before continuing.
• Schema: schema_version: 2. See references/curriculum-schema.md. If schema_version is missing or unknown, refuse and tell the user to update the skill.
• Failure: on 404 or network error, tell the user, suggest refresh curriculum, and offer to accept the meeting topic in one sentence as manual input. Do not fabricate content.

2. Meeting content

Each meeting in the manifest has two fields the skill uses:

• summary_md — the already-synthesized record of what was covered in that meeting.
• has_content: boolean — true when summary_md is real content, false when it's a pre-event stub.

If has_content is true: use summary_md as the source of truth. Do not fetch a separate transcript. Do not invent content.

If has_content is false (meeting hasn't happened yet): see "Handling future meetings" below.

Optional per-meeting endpoint (for direct drill-in without re-fetching the cohort list):

https://agency-lab.glebkalinin.com/api/curriculum/<cohort>/<meeting>.json

3. Personal vault

Read from the vault (see lab-context skill):

• goals.md — what the user wants
• tools-learned.md — what's already familiar (recent entries matter more)
• projects.md — what they're working on (best landing pad for tasks)
• profile.md — level, primary language
• homework/ — existing assignments (to avoid duplicates)

If the vault is missing or empty, point to lab-context first. Don't invent personal context.

4. Sandbox context (optional)

Read <vault>/sandbox/ if it exists. These files provide a simulated work environment (clients, customers, stakeholders, projects) that grounds homework in realistic scenarios.

• Real projects beat sandbox. If projects.md has an active project that fits the meeting topic, use it as the task landing pad.
• If no real project fits (or projects.md is empty), use sandbox entities to make the task concrete — reference specific clients, customers, or team members.
• Never mix sandbox and real-project context in the same task.
• Starter sandbox templates are bundled at <skill-dir>/sandboxes/{company,consultant,freelancer,solopreneur}/. If the user has no <vault>/sandbox/ and asks for one, offer to copy one of these.

Workflow

1. Resolve cohort — read <vault>/.config.md for cohort:. If missing, ask via AskUserQuestion (default claude-code-lab-05) and save it.
2. Fetch or reuse the curriculum manifest (see caching rules above).
3. Ask which meeting via AskUserQuestion, populated from manifest.meetings[] (number + title). Mark stubs visibly — append " (upcoming)" or " (not yet covered)" to titles where has_content is false. Single-select.
4. Ask parameters via a second AskUserQuestion:
   • "How much time do you have?" — 30 min / 1 hour / evening / weekend
   • "Where should this land?" — active project / playground / either
5. Read vault + optional sandbox (profile, goals, tools-learned, projects, existing homework, sandbox/).
6. Generate the assignment using the pedagogical rubric below, grounded in the selected meeting's summary_md.
   • If has_content: false, follow "Handling future meetings" below instead of improvising from the stub.
7. Offer recommended agents (optional step, silent when not applicable).
   • If the manifest has a top-level agents[] array AND the selected meeting has recommended_agents: look each up, tell the user "[Agent Name] could help — install to .claude/agents/?", confirm before installing.
   • Otherwise: check the bundled <skill-dir>/agents/ directory. If bundled agents are topic-relevant (e.g. prompt-coach for prompting meetings), offer to install from local copies. Confirm first.
   • Write confirmed agents to .claude/agents/<id>.md. Skip any that already exist.
   • Never install automatically. If the user declines, move on and don't re-offer.
8. Write the homework file to <vault>/homework/YYYY-MM-DD-meeting-<NN>-<slug>.md.
9. Update <vault>/progress.md:
   • If the file doesn't exist, create it from <skill-dir>/references/progress-template.md (substitute {{date}} with today's ISO date).
   • Find the section for this meeting (e.g. ## Meeting 03 — Prompt Engineering). If missing, create it.
   • Append a line: - [ ] Homework: [[homework/YYYY-MM-DD-meeting-NN-slug.md]].
   • Update the updated: date in frontmatter and bump homework_completed by 0 (increment only when the user marks it done).
   • Use Edit, never overwrite the file wholesale.
10. Summarise briefly in the chat: one sentence about the task + the reflection question. Don't dump the whole file.

Handling future meetings

When the selected meeting has has_content: false, the manifest's summary_md is a stub and can't ground the task. Options, in order of preference:

1. Use the title + description. The frontmatter usually encodes intent (e.g. Meeting 05: Subagents & Research (theory + demos)). If that's enough to design a preparatory task, tell the user: "This meeting hasn't happened yet. I'll design a prep task based on the planned topic (). Confirm before I write."
2. Ask the user to describe the planned topic in one sentence if the title is too vague.
3. Redirect to a completed meeting if the participant is unsure. Show only has_content: true meetings and ask again.

Never silently generate a task using fabricated content.

Pedagogical rubric

• An artifact is mandatory. The participant must end up with a file / folder / script / commit they can show at the next meeting. "Read the docs" ≠ a task.
• Open-ended, not a test. Frame as "build X that does Y" — the how is the participant's call.
• 80% familiar + 20% new. Most of the task leans on tools-learned.md. Exactly one new element — not two.
• Ground it in the summary. Pull specific topics / demos from summary_md. The task should feel like "continue what we did on screen".
• Land it in a real project when possible. Real project > sandbox > playground.
• Provocation, not instruction. "Figure out how the agent could verify its own work" > "create a hook that runs pytest after every Edit".
• One reflection question that can't be googled — only answered from experience.
• Smaller than feels right. 45 minutes finished > 3 hours abandoned.

Output format

Save to <vault>/homework/YYYY-MM-DD-meeting-<NN>-<slug>.md:

---
date: <YYYY-MM-DD>
meeting: <NN>
meeting_title: <title from manifest>
estimated_time: <30m | 1h | evening | weekend>
tools: [<tool1>, <tool2>]
related_goal: <quote from goals.md>
related_project: <[[projects#name]] or playground or sandbox/<entity>>
status: new
---

# <Task name>

## Why this matters for you
<1-2 sentences linking the task to the user's goal>

## Context from the meeting
<1-2 sentences summarising the relevant slice of `summary_md` — what the presenter showed>

## What you'll build
<clear description of the target — build, fix, or investigate>

## The new thing
<the single 20% — explicitly named>

## Done when
- [ ] <concrete observable result>
- [ ] <another one>

## Reflection
<one question worth asking yourself after you finish>

Rules

• One task per call. If the user wants more — suggest picking one.
• Never duplicate past homework. Check homework/ first. Continue an existing task as "part 2" or pick a different angle.
• Concrete over abstract. Not "practice MCP" but "write an MCP server with one tool get_current_weather, wire it into Claude Code, call it three times".
• Never create files outside the vault. For scratch dirs use <vault>/homework/<slug>/.
• If the manifest fetch fails, tell the user and offer refresh curriculum or accept a one-sentence topic. Don't fabricate.
• If the cache is stale (older than 7 days), warn and offer to refresh before generating.