/app-store-review-arbitrage

name: app-store-review-arbitrage
description: "Fetches low-star App Store and Google Play reviews, clusters them into broken-promise patterns, and generates a ranked copy brief with positioning opportunities."
version: 1.0.0

app-store-review-arbitrage

Convert a competitor's App Store or Google Play URL into a one-session GTM brief: ranked complaint clusters, a broken promise map, landing page headlines, and ad copy directions — all sourced from verbatim reviews.

---

Critical Rules (read before Step 1)

These rules apply throughout all steps. Violating any of them fails Self-QA (Step 6).

1. **Every quote must be verbatim.** No paraphrase, no grammar correction, no cleaning. Exact reviewer words only. 2. **No fabricated statistics.** Do not write "40% faster" or "2× more reliable" unless a reviewer explicitly used similar language. The Self-QA step checks for uncited percentages. 3. **Cluster names must use reviewer language.** Study the anti-pattern table in Step 3. 4. **Every headline and ad copy direction must cite its source cluster.** Format: `[cluster: "cluster-name"]`. 5. **Section 2 is always present** in the output — even when degraded. Never skip or omit it. 6. **No banned words** in any generated copy: powerful, robust, seamless, innovative, game-changing, streamline, leverage, revolutionize, transform.

---

Step 1 — Parse Input and Detect Platform

Accept a natural language prompt containing one app URL. Extract the URL.

**Platform detection:**

• `apps.apple.com` → App Store
• `play.google.com/store/apps/details?id=` → Google Play
• Any other URL → stop and respond: "Please provide a direct App Store or Google Play URL. I can't analyse review data from other sources."

**ID extraction (do this before calling the script):**

| Platform | What to extract | How | |---|---|---| | App Store | Numeric `app_id` | Digits after `/id` in the URL | | App Store | `country` | 2-letter code after `apps.apple.com/` (e.g., `us`, `gb`) | | Google Play | `package_name` | Value of `id=` query parameter |

Persist the extracted values — you will need them for the output filename in Step 7.

If `product_context` was provided in the user's prompt (what their own product does), store it — used to personalise copy in Step 5.

---

Step 2 — Collect Reviews & Metadata

Run the full fetch script:

python3 scripts/fetch_reviews.py "{app_url}" --output {tmpdir}/asr-raw.json

*(Note: Replace `{tmpdir}` with your operating system's temp directory, e.g., `/tmp` on macOS/Linux or `C:\Temp` on Windows).*

This fetches both the store description metadata and the reviews.

• **App Store:** iTunes API — free, no auth. App Store reviews are fetched via Apple's public iTunes RSS feed. Some apps return 0 reviews due to Apple's API limitations — in that case the skill continues with available data and logs a warning. Google Play is the primary supported path.
• **Google Play:** `google-play-scraper` package — free, no auth

If the script fails, read the error from stderr. Common causes:

• Package not installed: run `pip install google-play-scraper`
• App not found: verify the URL is a current, live listing
• Google Play API error: run `pip install --upgrade google-play-scraper` and retry

The script will print collection progress to stderr. Wait for it to complete. After completion, read `{tmpdir}/asr-raw.json` and display the collection summary to the user:

✓ Collected [N] reviews ([N] low-star 1–3★) from [platform]
  Date range: [oldest] to [newest]
  Package: [iTunes API | google-play-scraper]

**Check the exit code:**

• Exit 0 → collection succeeded, check `metadata.store_description`. If null: note this — Section 2 will use the degraded state. Proceed to Step 3.
• Exit 1 → error (read stderr message, surface it to user, stop)
• Exit 2 → **Gate 1 triggered** (< 10 low-star reviews found)

**Gate 1 — Low signal stop:** If the script exits with code 2, read the `gate_message` from `{tmpdir}/asr-raw.json` and surface it to the user verbatim. Do not proceed to Step 3. Do not produce a partial brief.

---

Step 3 — Complaint Clustering

Load `low_star_reviews` from `{tmpdir}/asr-raw.json`.

Cluster all low-star reviews into **4–6 named complaint themes.** Apply this formula to score each review:

complaint_weight = (4 - rating) × recency_factor

recency_factor:
  review age ≤ 90 days  → 1.0
  review age 91–365 days → 0.7
  review age > 365 days  → 0.4

`review age` = (today's date) − (review `date` field) in days.

`cluster_score` = sum of `complaint_weight` for all reviews in the cluster.

**Cluster naming — critical rule:**

You will want to write abstract names. Resist. Use the exact verb and noun from reviews.

| ❌ Abstracted (wrong) | ✅ Reviewer language (correct) | |---|---| | "Stability issues" | "Crashes when exporting to PDF" | | "Sync problems" | "Data lost after sync between phone and desktop" | | "Monetisation friction" | "Paywall appears after 3 days, not 14 as promised" | | "Performance degradation" | "App freezes every time I search" | | "Onboarding issues" | "Can't figure out how to invite a teammate" |

**Rules:**

• Each review belongs to exactly one cluster (assign to its dominant theme)
• Discard any cluster with fewer than 3 reviews — log it as noise
• Select 3–4 verbatim quotes per cluster: lowest star rating first, then most recent

**Gate 2 — Minimum cluster size:** After discarding sub-3-review clusters, check how many clusters remain.

**Gate 3 — Low-confidence flag:** If fewer than 3 clusters remain:

• Do NOT stop. Continue to output.
• Prepend this to the brief header immediately after the app metadata:

> ⚠ **LOW CONFIDENCE:** Only [N] complaint cluster(s) met the minimum evidence threshold (≥ 3 supporting reviews). Output reflects limited data. Consider a competitor with more reviews, or broaden the rating filter.

• Include Medium-tier clusters in the output (score ≥ 5)

**Tier classification** (for the leaderboard table in Section 1):