When you ask "how do I publish to npm" or "recipe for OAuth setup" or "wie mache ich X", the Academy probably has a recipe for it. Fire academy_get_recipe automatically:

{
  "UserPromptSubmit": [
    {
      "hooks": [
        {
          "type": "command",
          "command": "/home/simple/.claude/hooks/academy-recipe-trigger.sh",
          "timeout": 3
        }
      ]
    }
  ]
}

Where academy-recipe-trigger.sh:

#!/usr/bin/env bash
input=$(cat)
prompt=$(echo "$input" | jq -r '.prompt // .user_prompt // ""')
# Trigger phrases (DE + EN) that indicate the user wants step-by-step guidance
if echo "$prompt" | grep -qiE 'how (do|to)|recipe for|wie mache ich|wie geht|step.by.step|schritt für schritt'; then
  echo '{"hookSpecificOutput":{"additionalContext":"User asked for how-to guidance. Suggest calling academy_concept_search({query: \"<topic>\"}) to find related concepts, then academy_list_recipes({phase: <relevant>}) to find a step-by-step recipe. If the user already has a recipe slug, call academy_start_recipe({slug}) directly."}}'
fi

Why bash + context-injection instead of mcp_tool calling academy_concept_search? Because we need to extract the topic from the prompt and bash regex is too fragile for that. The injected context tells the assistant exactly which two tools to call with the right parameters.

Every session should start knowing where you stand: XP, rank, current league, next lesson. Wire it:

{
  "SessionStart": [
    {
      "hooks": [
        {
          "type": "mcp_tool",
          "server": "academy",
          "tool": "academy_stats",
          "input": {},
          "timeout": 10,
          "statusMessage": "Academy: loading stats..."
        },
        {
          "type": "mcp_tool",
          "server": "academy",
          "tool": "academy_next_lesson",
          "input": {},
          "timeout": 10,
          "statusMessage": "Academy: finding next lesson..."
        }
      ]
    }
  ]
}

Two parallel hooks. Both read-only. Both under 500ms. Result: every Claude Code session you start, the assistant already knows your XP, rank, streak, and what to recommend studying next. Without typing a command.

• Idempotent. academy_stats and academy_next_lesson are pure reads. academy_concept_search is read-only.
• Fast. All three under 500ms.
• Deterministic. Same time, same stats. Same query, same concept ranking.
• Side-effect-free without user trigger. All read-only.
• GDPR. Academy data flows from your Claude Code session to your own Academy tenant on studiomeyer.academy. No third-party sharing. Your XP and lesson progress are not exposed in any leaderboard unless you opt-in.

Advanced pattern: when you call academy_progress_complete for a lesson, automatically offer the quiz:

{
  "PostToolUse": [
    {
      "matcher": "mcp__academy__academy_progress_complete",
      "hooks": [
        {
          "type": "mcp_tool",
          "server": "academy",
          "tool": "academy_quiz",
          "input": { "lesson_slug": "${tool_input.lesson_slug}" },
          "timeout": 10,
          "statusMessage": "Academy: loading quiz..."
        }
      ]
    }
  ]
}

This pattern cascades: lesson-complete → quiz-fetch. Claude Code presents the quiz inline, user answers, you call academy_quiz_submit. Closes the loop without any "and now check the quiz" prompt.

# Edit settings.json with the SessionStart hooks above
claude

You should see both status messages flash on session start: "Academy: loading stats..." then "Academy: finding next lesson...". The assistant's first response should reference your current XP and the recommended next lesson without you asking.

# In the running session:
# "wie mache ich npm publish"

Status: nothing (UserPromptSubmit triggers context-injection only, not a tool call). But the assistant's response should now suggest specific Academy recipes, usually 9.1-npm-publish.

Academy has 23 tools. We only wired three. The rest (review, certificates, tutor, leaderboard) are user-driven explicit calls, not lifecycle-triggered. Hooks are for things the assistant SHOULD do automatically. Tools that need user-input (e.g. submitting a quiz answer) stay user-explicit.

If you do a lot of Academy work, you might also wire:

• Stop → academy_review to log session retrospective
• SubagentStop → academy_tutor to ask the tutor for follow-up suggestions

But those are personal preferences. Start with the 2-3 from this recipe.

Done

Five SaaS-MCPs (Memory, CRM, GEO, Crew, Academy) all have working hook bundles. Your Claude Code workflow is now fully agentic, the assistant uses MCP tools without you having to remember to call them.

Next: distribute these as plugins

The five bundles are ready to be packaged as Claude Code plugins on the studiomeyer-io/studiomeyer-marketplace. One claude plugin install studiomeyer-memory-hooks and the user gets all four Memory hooks at once. That packaging is recipe 9.6 (coming soon as part of Phase 9 Distribution).

Source

• Claude Code Hooks Reference
• Academy MCP, npm package
• studiomeyer.academy