reply-to-every-review

Drafts a reply to every unanswered App Store and Google Play review of the user's own app through App Store Connect / Play Console - each reply is a dry run with AppGoat guardrail scores (policy, tone, claims), and only approved replies are posted.

Finds the reviews of the user's app that have no developer reply, writes a short, specific reply to each, and posts the ones the user approves. Nothing is posted without a dry run the user has seen: every reply goes through propose_action (dry run, free) and then apply_action.

When to use

• The user owns the app and has connected App Store Connect and/or Play Console in AppGoat (https://dash.appgoat.ai/connections).
• Not for competitors' apps (use app-audit-lite) or for filing bugs (use reviews-to-github-issues).

Inputs

• app (optional): a linked app (App Store id or Play package). Default: every app linked to an asc or play_console connection; ask which one when there are more than three.
• since (optional): ISO date; default all unanswered reviews (Play only returns the last 7 days).

Ask the user once, before drafting, for anything you cannot infer: the app's voice (default: friendly, plain, first person plural "we"), the support address to point to (default: the support URL in the listing), and whether there is a known fix or release to mention.

Limits per run unless the user says otherwise: at most 30 replies, lowest stars and newest first.

Steps

Costs in AppGoat credits. Without the MCP, use the REST equivalents in reference/api.md.

1. list_connections {} - 0. Keep asc and play_console connections with status: active and the app linked. mode: read_only can dry-run but not post: tell the user and continue with drafts only. No connection: send the user to https://dash.appgoat.ai/connections and stop.
2. For each connection and app: read_connection {id, what: "reviews", app, unanswered: true, limit: 100} - 1 per page, follow next_cursor up to 3 pages. Each review has a review_id, rating, title, body, language, version and date. Drop reviews older than since.
3. Optional prioritising: get_reviews {app, rating: "1-3", intent: "bug,crash", limit: 100} - 1 returns AppGoat's AI tags (intent, severity, topic) for public reviews. Match by text and date; answer bug/crash reviews with severity ≥ 2 first and group same-topic reviews so their replies are consistent (not identical).
4. Draft one reply per review (no calls), following the decision rules. Reply in the review's language.
5. For each draft: propose_action {type: "review.reply", connection: , params: {app, review_id, text}, playbook: "reply-to-every-review"} - 0. It returns the action id, the before → after diff (an existing reply shows as before) and the guardrail result. A BLOCKS check (e.g. over 350 characters on Play) means: shorten or fix and propose again.
6. Show the user one table of all drafts (see Output format), flagged ones first, and ask which to post: "all", "all except 3, 7", or edits ("4: mention the export fix"). Re-propose edited ones.
7. For each approved row: apply_action {id} - 1 when posted. Outcomes:
   • applied: posted (App Store replies appear after Apple's review, usually within 24 h).
   • pending_approval: the org requires a person to approve in the dashboard; collect the approve_urls and give them to the user (one link per action, or https://dash.appgoat.ai/approvals).
   • "needs approval by a person" (403): same, hand over the link, do not retry.
   • "account changed" (409): someone replied meanwhile; show the new diff and skip unless the user wants to replace it (force: true).
8. reject_action {id, reason} - 0 for drafts the user declined, so they leave the approvals list.
9. Report what was posted, what waits for approval and what was skipped.

Budget: 1–3 credits to read + 1 per posted reply (dry runs and pending approvals are free).

Decision rules

• Answer the review, not a template. Reference one or two concrete points the user made (the feature, the bug, the version). Never copy the same reply across reviews.
• Length: 2–4 sentences. Play replies ≤ 350 characters (hard limit); App Store ≤ 600 as a house rule (limit 5970). Count before proposing.
• Negative (1–2★): thank, acknowledge the specific problem without arguing, give one concrete next step (update to the version with the fix if the user confirmed one, a setting, or the support address). Ask for details only through the support channel.
• Neutral (3★): thank, answer the question or suggestion, say if it is on the roadmap only if the user confirmed it.
• Positive (4–5★): thank them for something specific; one sentence is fine. No upsell.
• Never: promise dates, refunds, features or fixes the user did not confirm; offer incentives for changing the rating (against both stores' policies); ask for a better rating; include personal data (names, emails, order ids) of the reviewer or staff; blame the user or Apple/Google; link anywhere but the app's own support or privacy pages.
• Skip (do not reply, list them): spam, abusive reviews, reviews that are only a star with no text on Play, and legal threats (flag those to the user).
• Guardrail: flagged or any check under 0.7 → rewrite once before showing it; if still flagged, show it with the flag and let the user decide. Never apply a draft whose guardrail shows BLOCKS.
• Existing replies: this playbook only answers unanswered reviews; replacing a reply needs the user to ask for it.

Output format

Draft table (before step 7):

markdown

# Review replies: <app> (<store>) - <n> unanswered, <k> drafted

| # | ★ | Review (excerpt) | Reply draft | Chars | Guardrail | Action |
|---|---|---|---|---|---|---|
| 1 | 1 | "Login loops forever since 2.4…" | "Sorry about the login loop since 2.4…" | 212 | flagged: claims 0.62 (promises a fix) | act_… |
| 2 | 5 | "Love the widgets" | "Thank you! The widgets are…" | 96 | passed 0.95 | act_… |

Skipped: 2 spam, 1 legal threat (#14, please read it yourself).
Reply "post all", "post all except 1", or edit a row.

Run report (after step 7):

markdown

Posted 18 replies (App Store replies go live after Apple's review).
Waiting for approval in the dashboard (4): https://dash.appgoat.ai/approvals
Skipped 3, rejected 2 drafts. Credits used: <sum>.

Automate it

Daily drafts that a person approves in the AppGoat dashboard. The run calls apply_action once per draft, which submits it for approval (or posts it when the org has a Studio auto-apply rule for review.reply that the guardrail score clears). Without such a rule nothing is posted until someone approves it at https://dash.appgoat.ai/approvals.

GitHub Action (.github/workflows/appgoat-review-replies.yml, secrets ANTHROPIC_API_KEY, APPGOAT_API_KEY):

yaml

name: AppGoat review replies
on:
  schedule:
    - cron: "0 7 * * *"
  workflow_dispatch:
permissions:
  contents: read
jobs:
  replies:
    runs-on: ubuntu-latest
    steps:
      - name: Install AppGoat playbooks and MCP config
        env:
          APPGOAT_API_KEY: ${{ secrets.APPGOAT_API_KEY }}
        run: |
          git clone --depth 1 https://github.com/appgoat/skills /tmp/appgoat-skills
          mkdir -p .claude/skills
          cp -r /tmp/appgoat-skills/skills/* .claude/skills/
          # Hosted AppGoat MCP server; the key goes into a file, not the command line.
          jq -n --arg key "$APPGOAT_API_KEY" \
            '{mcpServers: {appgoat: {type: "http", url: "https://api.appgoat.ai/mcp", headers: {"x-api-key": $key}}}}' \
            > "$RUNNER_TEMP/appgoat-mcp.json"
      - uses: anthropics/claude-code-action@v1
        with:
          anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
          prompt: |
            Use the reply-to-every-review skill for all linked apps. This is an unattended run: use a friendly,
            plain voice and the support URL from the listing, promise nothing. Instead of asking the user in step 6,
            call apply_action once for every draft whose guardrail passed (this submits it for approval in the
            dashboard), and reject_action the rest with the reason. Never call apply_action with force.
          claude_args: |
            --mcp-config ${{ runner.temp }}/appgoat-mcp.json
            --allowedTools mcp__appgoat__list_connections,mcp__appgoat__read_connection,mcp__appgoat__get_reviews,mcp__appgoat__propose_action,mcp__appgoat__apply_action,mcp__appgoat__reject_action,mcp__appgoat__get_action

Claude Code routine:

text

/schedule weekdays 09:00 - Run the reply-to-every-review playbook for my linked apps. Draft replies, submit the ones whose guardrail passed for approval (apply_action once each), and send me the approvals link with a one-line summary per app.