launch-summary

Instructions

Generate a non-developer-friendly launch summary from merged pull requests across four GitHub repos: Atllas-Inc/codebase, Atllas-Inc/aicc-queues, Atllas-Inc/neema-simple-hyzl, and Atllas-Inc/pointsgpt.

The skill takes one argument, the window: daily or weekly. If the user gives no argument, use daily. The window sets the timeframe and nothing else. Every later step is shared.

Step 1: Determine the date range

• daily: the last 24 hours, a rolling window ending now.
• weekly: the current calendar week, Monday 00:00 UTC through today.

If the user specified a different range, a specific day, or a specific week, use that instead.

Step 2: Fetch merged PRs from every repo

Only PRs merged into the repo's default branch count. The base: filter excludes PRs merged into release branches and feature branches.

Never hardcode base:master. The default branch differs per repo: codebase and aicc-queues use master, neema-simple-hyzl and pointsgpt use main. Read it from repos/<repo> so a repo added later needs no edit here.

Never use gh pr list. It calls the GitHub GraphQL API, which Claude Code sessions block with a 403. Use the REST endpoints below.

search/issues filters on merge date server-side, so no --limit can drop a PR that was opened long ago and merged inside the window. It returns the author, but not the merger and not the changed files. Step 2 therefore calls pulls/{n} for every PR to read merged_by. It also calls pulls/{n}/files for each repo that ships a phone app, to tell Mobile PRs from App PRs.

Two repos ship one: codebase under apps/atllas-app/, and pointsgpt under ios/. The mobile_prefix function below holds the prefix per repo. An empty value means no phone app.

product_name holds the customer-facing name per repo. The summary is grouped by that name, so every repo needs a row.

display_name turns a GitHub login into a person's real name, and caches each lookup in /tmp/ls_names.tsv so one person costs one call.

Add a repo by appending it to the for REPO in ... list and giving it a product_name row. Never put the repo list in a variable. This block runs under zsh, which does not word-split an unquoted variable, so a list in a variable is read as one repo name.

Run the whole block in one Bash invocation so SINCE is computed once:

set -euo pipefail
WINDOW=daily   # daily | weekly

case "$WINDOW" in
  daily)
    SINCE=$(date -u -v-24H +%Y-%m-%dT%H:%M:%SZ 2>/dev/null || date -u -d "24 hours ago" +%Y-%m-%dT%H:%M:%SZ 2>/dev/null)
    [ -n "$SINCE" ] || { echo "ERROR: could not compute a 24-hours-ago timestamp with this system's date(1). Aborting." >&2; exit 1; }
    HEADER="## 🚀 Daily Launch Summary: $(date -u +%Y-%m-%d)"
    EMPTY="Nothing shipped in the last 24 hours."
    ;;
  weekly)
    WEEK_START=$(date -v-Mon +%Y-%m-%d 2>/dev/null || date -d "last Monday" +%Y-%m-%d 2>/dev/null)
    [ -n "$WEEK_START" ] || { echo "ERROR: could not compute the start of the week with this system's date(1). Aborting." >&2; exit 1; }
    SINCE="${WEEK_START}T00:00:00Z"
    HEADER="## 🚀 Weekly Launch Summary: Week of ${WEEK_START}"
    EMPTY="Nothing shipped this week."
    ;;
esac
UNTIL=""
NAMES=/tmp/ls_names.tsv
: > "$NAMES"
echo "Window starts: $SINCE"
echo "Header: $HEADER"
echo "Empty-result line: $EMPTY"

mobile_prefix() {
  case "$1" in
    codebase) echo '^apps/atllas-app/' ;;
    pointsgpt) echo '^ios/' ;;
    *) echo '' ;;
  esac
}

product_name() {
  case "$1" in
    codebase) echo 'Atllas' ;;
    aicc-queues) echo 'AI Calling' ;;
    neema-simple-hyzl) echo 'Hyzl' ;;
    pointsgpt) echo 'PointsGPT' ;;
    *) echo "$1" ;;
  esac
}

display_name() {
  [ -n "$1" ] || { echo ''; return; }
  local hit
  hit=$(awk -F'\t' -v l="$1" '$1==l {print $2; exit}' "$NAMES")
  if [ -z "$hit" ]; then
    hit=$(gh api "users/$1" --jq '.name // .login')
    printf '%s\t%s\n' "$1" "$hit" >> "$NAMES"
  fi
  echo "$hit"
}

search_prs() {
  gh api -X GET search/issues \
    -f q="repo:$1 is:pr is:merged base:$2 merged:>=$SINCE" \
    -f per_page=100 --paginate \
    --jq '.items[] | {number, title, mergedAt: .pull_request.merged_at, body, author: .user.login, labels: [.labels[].name]}'
}

: > /tmp/ls_prs.jsonl
for REPO in Atllas-Inc/codebase Atllas-Inc/aicc-queues Atllas-Inc/neema-simple-hyzl Atllas-Inc/pointsgpt; do
  NAME="${REPO#*/}"
  BASE=$(gh api "repos/$REPO" --jq .default_branch)
  PRODUCT=$(product_name "$NAME")
  PREFIX=$(mobile_prefix "$NAME")
  echo "$REPO: base $BASE, product $PRODUCT"
  search_prs "$REPO" "$BASE" | while read -r row; do
    n=$(jq -r .number <<<"$row")
    m=false
    if [ -n "$PREFIX" ]; then
      if gh api "repos/$REPO/pulls/$n/files?per_page=100" --paginate --jq '.[].filename' \
           | grep -q "$PREFIX"; then m=true; fi
    fi
    MERGER_LOGIN=$(gh api "repos/$REPO/pulls/$n" --jq '.merged_by.login // ""')
    AUTHOR=$(display_name "$(jq -r .author <<<"$row")")
    MERGER=$(display_name "$MERGER_LOGIN")
    jq -c --arg repo "$NAME" --arg product "$PRODUCT" --argjson m "$m" \
          --arg authorName "$AUTHOR" --arg mergerName "$MERGER" \
      '. + {repo: $repo, product: $product, mobile: $m, authorName: $authorName, mergerName: $mergerName}' <<<"$row"
  done >> /tmp/ls_prs.jsonl
done

jq -s --arg until "$UNTIL" '[.[] | select($until == "" or .mergedAt < $until)]' /tmp/ls_prs.jsonl

A repo with no mobile_prefix row gets mobile: false on every PR (App).

An empty mergerName means the merger is no longer readable, for example a deleted account. Print the author alone for that PR.

Bounded window. The search filter is one-sided, so "what shipped yesterday" also returns everything merged today. To bound the far end, set UNTIL to a timestamp in the same format. UNTIL is empty by default, and an empty value keeps every PR.

Step 3: Analyze and categorize

For each PR, read the title and the Description and Changes sections of the body. Write plain-English summaries. Never copy developer jargon or ticket IDs into the output.

Exclude these from the summary:

• PRs that only touch CI/CD pipelines, linting configs, test infrastructure, or dev tooling with no user impact
• Dependency bumps or version updates with no feature changes
• Code refactors or internal restructuring with no user-visible effect
• MCP server config changes, developer debugging tools

Include everything that affects what users see, experience, or can do. Include a fix even when the thing it fixed was itself broken by a recent change.

Group every included PR three levels deep, in this order:

1. Product: the product field. Order the products as the for REPO in ... list orders them: Atllas, AI Calling, Hyzl, PointsGPT.
2. Mobile, then App, inside each product. Mobile: PRs where mobile: true. That is apps/atllas-app/ in codebase (the React Native/Expo app) and ios/ in pointsgpt (the iPhone app). App: everything else, which is the web dashboards, admin sites, and servers of that product.
3. Category, inside each Mobile or App block.

A PR that touches both its repo's phone-app prefix and backend paths gets mobile: true, so it goes under Mobile only. If its backend half has user impact of its own, add a second bullet for that impact under App, in the same product.

Use these categories (only include a category if it has items):

1. New Features: brand-new capabilities that did not exist before
2. Improvements: enhancements to existing features (faster, smarter, more accurate, better UX)
3. Bug Fixes: things that were broken and are now fixed
4. Reliability & Performance: backend changes that affect stability, speed, or uptime, even if invisible to the user

Step 4: Write the summary

The first line is the HEADER string from Step 2, printed verbatim. All dates in the output use YYYY-MM-DD format.

[HEADER]

### [Product]

#### 📱 Mobile

**New Features**
- **[Short feature name]**: [Half a sentence max. What can users now do?] ([Author], merged by [Merger])

**Improvements**
- **[Short name]**: [Half a sentence max. What got better?] ([Author], merged by [Merger])

#### 💻 App

**Bug Fixes**
- **[Short name]**: [Half a sentence max. What was broken?] ([Author], merged by [Merger])

**Reliability & Performance**
- **[Short name]**: [Half a sentence max. What got more stable?] ([Author])

---
_[N] pull requests merged_

Repeat the whole product block for each product that has included PRs. Omit any product, any Mobile or App block, and any category with no items. Print the #### 📱 Mobile and #### 💻 App headings only when the product has both blocks. A product with one block prints its categories directly under the product heading.

If no PRs merged in the window, print the EMPTY string from Step 2 and nothing else.

Attribution rules:

• The credit goes in parentheses at the end of the bullet, after the period.
• Use the first name from authorName and mergerName. Add the last initial for two people who share a first name, for example Simon K.
• Write (Phillip, merged by David). Write (Phillip) alone when the author merged their own PR, or when mergerName is empty.
• A bullet that groups several PRs lists each distinct author, then the mergers: (Phillip and Simon, merged by David).
• An empty authorName means a bot opened the PR. Drop the whole credit for that bullet.

Tone guidelines:

• Ultra-concise: each bullet is a fragment, not a full sentence. Think changelog entry, not explanation.
• No filler words: drop "now", "previously", "instead", "in order to". Just the fact.
• "AI calling" not "AICC", "contacts" not "recipients", "dashboard" not "portal"
• Never use a technical term: no Firestore, Redis, UUID, cron, CSV (say "spreadsheet"), API, etc.
• Group closely related PRs into a single bullet, only when they share a product and a Mobile or App block. Never merge a Mobile PR and an App PR into one bullet, and never merge PRs from two products.
• Keep the bold name to 1 to 3 words maximum

Print the formatted summary to the user. Do not save it to a file unless the user asks.