agentleFS
glebisglebis/claude-skills/lab-retro/SKILL.md

lab-retro

Skill400 starsChanged 4 months agoMITView on GitHub

Final retrospective and self-assessment for participants of Claude Code Lab. Runs four sequential interactive parts — progress audit, best prompt, monthly plan, and feedback — using AskUserQuestion. Triggers on "/lab-retro", "lab retrospective", "claude code lab final", or after completing the 6-week Claude Code Lab cohort.

  • Sends data out

We split this skill into the rules it gives an agent

10 rules. About 0 are common, 6 are distinctive, and 4 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈6

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 4

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Default flow: run all four parts in order.

    line 12
  2. Very specific:

    All artifacts are saved into a single folder lab-retro-output/ in the current working directory:

    line 16
  3. Distinctive:

    Create the folder if missing.

    line 22
  4. Distinctive:

    If risk = high, insert a mandatory rest day into the plan and a recommendation to read the AI hygiene slide again.

    line 108
  5. Distinctive:

    Stop-conditions (when to pause)

    line 114
  6. Very specific:

    If the request fails, fall back to local files only and tell the participant: "submit failed — your feedback is saved locally in …

    line 152
  7. Distinctive:

    After all four parts, print:

    line 158
  8. Very specific:

    Always use AskUserQuestion for structured questions — don't ask in plain text

    line 166
  9. Distinctive:

    Never skip a part silently.

    line 167
  10. Distinctive:

    Don't be sycophantic in the feedback report — surface honest patterns

    line 169

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Claude Code Lab — Final Retrospective

This skill walks a Claude Code Lab graduate through four sequential exercises that consolidate their learning, capture their best work, plan next steps, and collect structured feedback for the organizer.

How to run

Default flow: run all four parts in order. The user can also jump to a specific part with /lab-retro 2 (or just say "part 3").

Between parts, briefly summarize what just happened and ask "ready for part N?" so the user controls the pace.

All artifacts are saved into a single folder lab-retro-output/ in the current working directory:

  • 01-progress.md
  • 02-best-prompt.md
  • 03-month-plan.md
  • 04-feedback.json + 04-feedback-report.md

Create the folder if missing.


Part 1 — Progress audit

Goal: help the participant see concrete before/after.

Use AskUserQuestion:

  1. "Что вы умели ДО лаборатории?" (multiSelect)
    • Работал с ChatGPT/Claude в браузере
    • Использовал CLI инструменты
    • Писал код
    • Автоматизировал задачи
    • Работал с API
  2. "Что вы умеете ПОСЛЕ?" (multiSelect)
    • Работаю с Claude Code ежедневно
    • Настроил MCP-серверы
    • Создал свой Skill
    • Автоматизировал реальную задачу
    • Задеплоил что-то в веб
    • Использую субагентов
    • Пишу и читаю CLAUDE.md осознанно
  3. "Сколько часов в неделю экономит Claude Code?" (singleSelect)
    • <2
    • 2–5
    • 5–10
    • 10+

Then output a markdown table comparing before/after with skill levels (0–5) and save to lab-retro-output/01-progress.md.


Part 2 — Best prompt

Goal: turn one prompt the participant is proud of into a reusable Skill.

Ask the participant: "Скопируйте или опишите ваш самый полезный промт из лабы."

Then AskUserQuestion:

  1. "Для какой задачи был промт?" (singleSelect)
    • Автоматизация рутины
    • Создание контента/документации
    • Анализ данных/исследование
    • Прототипирование/разработка
    • Личный workflow / PKM
  2. "Что сделало его эффективным?" (multiSelect)
    • Хороший контекст в CLAUDE.md
    • Чёткие критерии успеха
    • Разбиение на шаги
    • Использование Skills/MCP
    • Примеры в промте
    • Ограничения и анти-критерии

Reformat the prompt as a proper Skill (frontmatter + body), suggest an description line that would trigger it, and save to lab-retro-output/02-best-prompt.md. Suggest where to put it (~/.claude/skills/<name>/SKILL.md).


Part 3 — Month plan

Goal: concrete 4-week plan so momentum doesn't die after the cohort.

AskUserQuestion:

  1. "Главная рабочая задача на месяц?" (singleSelect)
    • Проект для клиента/работодателя
    • Свой продукт/стартап
    • Автоматизация текущих процессов
    • Обучение и развитие навыков
  2. "Сколько времени в неделю готовы уделять?" (singleSelect)
    • 15–30 минут
    • 1–2 часа
    • 3–5 часов
    • Каждый день
  3. "Какой риск выгорания вы оцениваете для себя?" (singleSelect)
    • Низкий — у меня устойчивый ритм
    • Средний — иногда залипаю
    • Высокий — уже узнал себя в красных флагах

If risk = high, insert a mandatory rest day into the plan and a recommendation to read the AI hygiene slide again.

Generate lab-retro-output/03-month-plan.md with:

  • 4 weekly goals
  • 1–3 concrete prompts per week
  • Success criterion per week
  • Stop-conditions (when to pause)

Part 4 — Feedback for the organizer

Goal: structured feedback that goes back to the lab organizer.

AskUserQuestion:

  1. "Оцените лабу в целом (NPS)" (singleSelect: 0–10)
  2. "Самая ценная встреча?" (singleSelect: M01 / M03 / M05 / M07 / M09 / M11)
  3. "Самая ценная тема за все 6 недель?" (multiSelect)
    • Основы Claude Code
    • Промтинг и контекст
    • Архитектура и субагенты
    • MCP / Skills / Hooks
    • Agent SDK и деплой
    • Evals и качество
    • AI-гигиена
  4. "Что улучшить?" (free text)
  5. "Главное препятствие, с которым вы столкнулись?" (free text)
  6. "Согласны ли поделиться отзывом публично?" (singleSelect: да / да-анонимно / нет)

Save TWO files:

  • lab-retro-output/04-feedback.json — structured for the organizer
  • lab-retro-output/04-feedback-report.md — human-readable summary for the participant

Then submit to the public proxy (no secrets needed):

curl -sS -X POST https://lab-feedback-proxy.vercel.app/api/feedback \
  -H "Content-Type: application/json" \
  -d "$(jq -nc --arg name "<participant name>" --slurpfile notes lab-retro-output/04-feedback.json '{name:$name, notes:($notes[0]|tostring)}')"

The proxy forwards to Baserow table 746002 with a server-side token. Response is {"ok":true,"row_id":<N>}. Confirm row ID with the participant.

If the request fails, fall back to local files only and tell the participant: "submit failed — your feedback is saved locally in lab-retro-output/04-feedback.json, send it to the organizer manually."


Closing

After all four parts, print:

  1. Where the four files live
  2. One sentence: "Ваш фреймворк — LOOP. Ваш план — в файле 03. Ваш следующий шаг — открыть его в понедельник утром."
  3. Suggest committing the lab-retro-output/ folder to a personal repo or vault.

Constraints

  • Always use AskUserQuestion for structured questions — don't ask in plain text
  • Never skip a part silently; if the user opts out, write a one-line stub in the corresponding file
  • Russian by default (audience is Russian-speaking); switch to English if the user replies in English
  • Don't be sycophantic in the feedback report — surface honest patterns

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.