A skill for Claude Code

Your week of code, back in your head.

debrief turns a week of commits, pull requests and AI sessions into a quiz and a dashboard, so your team keeps a grip on code it no longer writes by hand.

npx skills add arimet/debrief
The This week tab of debrief: the week's features as cards, each with its summary, its source and the cases handled and not handled.

When AI writes the code, you spend the week reading and approving. The pull requests pass, the features ship, and on Friday nobody can quite say which rule they follow, or why.

Reviewing what someone else produced is not the same as knowing it. debrief puts you back in the position of the one who produces: once a week, you say what shipped before the tool shows you.

Produce before you read

Three moves, ten minutes, once a week.

The order matters. You write first, from memory, then the page tells you what you missed. Try the second one below.

1

Retell the week from memory

The quiz opens on a blank box. What shipped this week, and for whom? A few sentences, in your own words, without looking.

Tell your week

Before the quiz, without reopening the week tab.

This week we added offline sync to the scorer's table, coaches got a subscription to edit the playbook, and plays can now be shared by link
SkipSave and start the quiz
2

Make the call the team made

Decision questions give you the context of a real choice from the week. You pick, you say what you rule out, and only then do you see what the team chose and why.

A different choice is not a wrong answer. It is the conversation you should have had before the merge.

The scorer's phone loses the network in the middle of a game. What should the app do?

3

See what you got wrong

Ask Claude to compare your retelling. Each claim is checked against what the pull requests, specs and commits say, and the features you never mentioned are listed.

The vocabulary check catches the quiet ones: a word that is not the one the team uses usually means the logic will land in the wrong place.

Against the pull requests

  • Drifts from the PRs
    we added offline sync to the scorer's table

    The opposite: offline mode was removed. Every action is written to the database first.

  • Vocabulary
    subscription

    The sources say "access level". There is no billing, coaches get a code.

  • Matches
    plays can now be shared by link

    By link, image, PDF or GIF.

One local page

Three tabs, one for each question.

Everyone sees the same week. The role only changes the questions and the part of the dashboard written for you.

This week

What shipped, in plain words: each feature with its source, the cases it handles and the ones it does not.

Business: what changes for users, and the questions to ask the client.

Technical: most-changed files, untested files, blind spots, commit heatmap.

Quiz

The retelling, then 5 to 10 questions picked from the work you did not write yourself first.

  • What happens when a user does this?
  • Where does this rule live?
  • Why was it built this way?
  • What would you have decided?

Review

Your retelling, compared claim by claim.

How the team worked: descriptions, approvals, tests, merge speed.

How you worked with Claude: framing, evidence, control, and who held the pen.

Under the hood

Facts first, then Claude.

Scripts measure, Claude writes, and the page tells you which is which. Every statement carries its source, and a rule read only in the code says so.

Measured by the scripts Written by Claude
  1. Extract

    no AI

    Commits, merged pull requests, risky files, blind spots, team signals, your redacted sessions.

  2. Summarize

    subagents

    Up to 8 in parallel, one card per feature, each statement tagged with its source.

  3. Assemble

    no AI

    Cards merged. Dependency, CI and release PRs listed without a summary.

  4. Quiz & review

    Claude

    Questions start with your blind spots and the risky files.

  5. Check & render

    no AI

    Every file, commit and id verified, then one self-contained HTML page.

About 5 minutes for a week of 100 commits and 30 pull requests. A rerun reuses the cards of unchanged pull requests.

tests_weakened

Assertions or test cases removed while the code changed: the tests rewritten until they pass.

ai_without_tests

A commit co-authored by an AI assistant that touches no test.

fast_large_merge

400 lines or more merged within 30 minutes of opening.

no_human_approval

Merged without a single human approval.

blind spots

Files changed this week whose history you never touched. They feed the quiz first.

who holds the pen

How many of your messages to Claude only said "ok" or "go".

Private by default

A mirror, not a dashboard for managers.

Stays on your machine

Served on 127.0.0.1, answers in a local file, .debrief/ kept out of git. Sharing is a download you choose to make.

Nobody is ranked

The quiz never asks who wrote what. The repository review speaks about the team, never about one person.

Sessions are redacted

Keys, tokens, .env values, emails and your home path are masked before Claude reads a session.

Install

One command, then /debrief every week.

Get the skill

npx skills add arimet/debrief

Or from a clone, so a git pull keeps it up to date:

git clone git@github.com:arimet/debrief.git
ln -s "$(pwd)/debrief/skills/debrief" \
  ~/.claude/skills/debrief

Run it

In a repository, in Claude Code:

/debrief/debrief --branch develop/debrief --lang encompare my retelling

Or in your own words: "quiz me on this week", "what did we ship this week".

  • Claude Code, git and Python 3.9+, standard library only. The skill checks Python first and explains how to install it.
  • gh for pull requests. Without it, the week is read from git alone. GitLab through glab is partial.