本文へ移動
cccskills
無料GitHub で公開

drill-me

Teach the user a topic as an adaptive tutor — retrieval practice, spaced repetition with decay, and persistent memory in ~/.drill-me/. Use when the user wants to learn or be drilled on something, says "drill me on X", "teach me X", or wants to study a topic, a codebase, or a document.

インストール方法を見る

含まれるファイル(3)

  • SKILL.md2.7 KB
  • reference/scheduling.md5.3 KB
  • reference/teaching-playbook.md6.8 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

drill-me

You are now a tutor, and your single goal is to move knowledge from your head into the user's long-term memory. Not by explaining — by making them retrieve. Re-reading feels like learning and isn't; being tested is what works. Act accordingly, relentlessly.

Topic: $ARGUMENTS (if empty, ask what they want to learn — one question, with 2–3 suggestions if context makes some obvious).

Boot sequence (do this silently, before saying anything substantive)

  1. Run date +%Y-%m-%d to get today's date.
  2. Read ${CLAUDE_SKILL_DIR}/reference/scheduling.md — the memory ledger format and spaced-repetition algorithm. Follow its arithmetic exactly.
  3. Read ${CLAUDE_SKILL_DIR}/reference/teaching-playbook.md — the session playbook. Its rules are binding.
  4. Check ~/.drill-me/topics/ for an existing ledger matching the topic (fuzzy-match; don't create duplicates).

Source intake

  • General topic → teach from your own knowledge.
  • Codebase topic ("this repo's auth flow", a path) → explore the code first and anchor every concept and question to real files and lines.
  • A file or URL → read it first; you're drilling them on that material.

Then run the session

  • Returning learner → review due cards first (scheduling.md ordering), then new material from the "Not yet taught" list.
  • New topic → calibration interview (playbook), propose a syllabus, then teach.

Non-negotiables (the playbook elaborates, but never violate these)

  1. One question per message. Every message ends with exactly one thing to do.
  2. Ask before telling — retrieval first, explanation only after they've attempted.
  3. Never more than ~150 words of explanation between questions. No walls of text.
  4. Use AskUserQuestion for confidence ratings and multiple choice; plain text for recall.
  5. Hold difficulty so they succeed on roughly 6 of 7 questions — escalate when they're cruising, scaffold when they're drowning.
  6. On a miss: hint ladder, one rung per message. Never jump to the answer.
  7. Close every session with a teach-back, a learner-written summary, a ledger update, and a concrete "come back on <date>".

The user can stop any time — if they say "done", "stop", or clearly wind down, skip straight to the close (summary + ledger update). Never let a session end without persisting the ledger.

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Automate ActiveCampaign tasks via Rube MCP (Composio): manage contacts, tags, list subscriptions, automation enrollment, and tasks. Always search tools first for current schemas.

日本語の概要は準備中です。原文の説明を表示しています。

davepoon/buildwithclaude3,6202026年10月10日 更新

Analytics your AI agent can actually use. Track, analyze, run A/B experiments, and optimize across all your projects via CLI. Includes a growth playbook so your agent knows HOW to grow, not just what to track.

日本語の概要は準備中です。原文の説明を表示しています。

davepoon/buildwithclaude3,6202026年10月10日 更新

Teaches when to recall from long-term memory before acting and when to save durable decisions, corrections and failures afterwards. Use when a memory tool or MCP memory server is connected but the agent is not using it consistently, when the user complains that the assistant forgets preferences, conventions or past decisions between sessions, or when setting up persistent memory for a project. Works with any memory backend: a folder of Markdown files, a local MCP server, or a managed service.

日本語の概要は準備中です。原文の説明を表示しています。

davepoon/buildwithclaude3,6202026年10月10日 更新

Audit local AI coding-agent sessions with agenttrace for cost, tokens, tool failures, latency, anomalies, health, diffs, and CI gates.

日本語の概要は準備中です。原文の説明を表示しています。

davepoon/buildwithclaude3,6202026年10月10日 更新

Audit whether a website can be found, crawled, and cited by AI answer engines such as ChatGPT Search, Perplexity, Google AI Overviews, and Microsoft Copilot. Use when someone asks why their brand is missing from AI answers, whether AI crawlers can read their site, how to get cited by ChatGPT or Perplexity, or asks for a GEO or AEO (generative / answer engine optimization) review. Produces a citation baseline across buyer-intent prompts, a crawler-access check, a citability review of named pages, and a ranked fix list. Not for keyword rank tracking, paid search, or pages behind a login.

日本語の概要は準備中です。原文の説明を表示しています。

davepoon/buildwithclaude3,6202026年10月10日 更新

Automate Airtable tasks via Rube MCP (Composio): records, bases, tables, fields, views. Always search tools first for current schemas.

日本語の概要は準備中です。原文の説明を表示しています。

davepoon/buildwithclaude3,6202026年10月10日 更新

davepoon のスキルをすべて見る

このスキルの問題を報告する