Qaf Customer Feedback Report
Date: 2026-08-24 Source: 702 Gmail threads (2026-03-23 to 2026-08-24), mostly post-V1-launch (2026-08-10)
TL;DR: The V1 launch wave surfaced two P0 bugs — Persian/Dari/Pashto/Kurdish text renders as broken, disconnected letters (12 reports, our single most-reported bug), and sign-in/OTP failures that fully lock users out (7 reports). Blind users are a large, vocal segment and the app is effectively unusable for them with screen readers (16 threads across bugs and feature requests). The lifetime offer generated the highest-volume thread type in the inbox (25 inquiries), but $1,000 is out of reach for many: 10 threads ask for a donation path or say they are waiting for a cheaper monthly plan — clear demand for a lower tier. Finally, trust is the recurring theme in both praise and complaints: citation grounding is our most-praised differentiator, while AI-written scripture errors and one-school framing are the most trust-damaging defects.
Inbox at a glance
| Metric | Value |
|---|---|
| Threads classified | 702 |
| Total classified items | 728 |
| Date range | 2026-03-23 to 2026-08-24 |
A thread that raises multiple issues (e.g. a bug plus a feature request) is counted once per category it touches, which is why items (728) exceed threads (702).
| Category | Items | |
|---|---|---|
| internal_system | 370 | automated |
| feature_request | 71 | human |
| promotional_marketing | 63 | automated |
| bug | 61 | human |
| question | 54 | human |
| membership_paid_plans | 40 | human |
| no_reply_needed | 33 | human |
| feedback | 21 | human |
| cold_outreach | 9 | human |
| automated_notification | 6 | automated |
Human vs automated: 289 items are human mail (bugs, feature requests, questions, membership, feedback, cold outreach, no-reply-needed) vs 439 automated/system items (internal system, promotional, notifications). The real customer signal is roughly 250 threads, almost all arriving after the 2026-08-10 launch.
Top bugs
P0: Persian/Dari/Pashto/Kurdish rendering — 12 reports
Twelve distinct users report that non-Arabic Arabic-script languages render as disconnected or separated letters, garbled fonts, missing glyphs (e.g. the letter "ye"), and wrong RTL character/word order — content is unreadable. The current Arabic font simply lacks the letters used by Pashto and Kurdish Sorani; users suggest a broader font such as Noto Naskh Arabic. One reporter is a channel owner relaying constant complaints from his audience, and one user confirmed the bug was still unfixed two weeks after reporting it. This is our most-reported bug and it blocks an entire language segment. Details: arabic-script-font-rendering.
P0: Sign-in and OTP failures — 7 reports
Seven users could not sign in or register: emailed verification codes are rejected as incorrect (one user sees the digits reversed and rejected either way), resent codes never arrive, both Google and email sign-in fail during registration, one user could not sign back in after logging out, and one was blocked by the temp-email anti-spam rule (migrated manually). Two more code-related threads sit in the questions pile (login-verification-code-issues), both unanswered. These failures fully block access, and users worry about losing chat history. Details: login-otp-failures.
P1: Screen-reader accessibility — 7 bug reports (plus 9 feature-request threads)
Seven blind users on iOS and Android report the app is incompatible with VoiceOver/TalkBack: most buttons (send, chat list, incognito, answer actions) are unlabeled, the VoiceOver cursor jumps around, conversations can be read as titles but not opened, and cited sources cannot be reached after results appear. Combined with the nine feature-request threads on the same topic, blind users are one of the loudest segments in the inbox — and the app is effectively closed to them. Details: screen-reader-accessibility.
P1: Answers fail to complete — 4 reports
The core Q&A flow intermittently fails to deliver: no response at all (including follow-ups), repeated connection-failure errors while searching, answers cutting off at unpredictable spots, and research mode stalling with the answer completing only minutes or hours later. This is a recurring reliability failure of the app's primary function. Details: answers-fail-to-complete.
P1: Pipeline artifacts leak into answers — 4 reports
Internal tokens are leaking into answer text: raw UUIDs mixed into answers instead of rendered source references, programming code appearing in fiqh answers, and follow-ups repeatedly failing with the model replying just "DONE" — including for one first-time user on his very first exchange. All point to answer post-processing failing to strip or render internal tokens and citation markup. Details: raw-code-in-answers.
P1: AI misquotes scripture — 2 reports
A detail-oriented user caught the AI miswriting ayah 36:57 in chat (tapping the verse showed the correct text) and, earlier, quoting a hadith with Arabic spelling mistakes and wrong i'rab. The underlying verse data is correct; the model-generated text is not. For an Islamic app this is a trust-critical defect regardless of report count. Details: ai-misquotes-scripture.
P1: Stripe checkout link errors on mobile — 1 report
A customer buying the lifetime offer got an error when tapping the Stripe link on his phone; it only worked via "Open in Browser". One report and a workaround exists, but it sits directly in the revenue path of the launch promotion. Details: stripe-checkout-link-error.
Top feature requests
P1: Partial text selection and copy — 10 reports
Ten users across iOS, Android, and web cannot select answer text; they must copy the entire answer via the icon or cannot copy at all. They want ChatGPT-style free selection; one also wants Word/PDF export with Arabic formatting preserved. Abdellatif has already told several of them partial copy is coming — this is now a promise to keep. Details: copy-select-text.
P1: Screen-reader accessibility — 9 threads
Mostly blind users asking for full TalkBack/VoiceOver/NVDA compatibility: labeled buttons and icons, navigable screens, and an in-app read-answers-aloud option. One user has offered to beta-test accessibility builds. Same underlying work as the P1 accessibility bug above; together they represent 16 threads. Details: screen-reader-accessibility.
P1: Expand the source library — 9 threads
Users ask Qaf to grow beyond Al-Shamela: the Al-Jame' platform (aljam3.com, 63k books, requested twice), the King Abdullah Sunnah complex, Fiqh Council declarations, IslamQA/IslamWeb, works of Ibn Baz, Ibn Uthaymeen, Ibn Taymiyyah, al-Albani and Ibn al-Qayyim, South Asian texts (Tafheem ul Quran), Allama Naser Sobhani's works, and Shia sources plus academic databases behind a research mode. Broader coverage of Islamic scholarship is repeatedly framed as the core value driver. Details: expand-source-library.
Feedback themes
Praise (16 of 21 feedback threads). A large launch wave of heartfelt thanks — including an Al-Azhar hadith professor, a Shariah scholar, an elderly Quran-circle teacher, and a software engineer. Users describe Qaf as "an indispensable study companion, teacher, and friend" that "never disappoints," and several actively promote it to research groups, WhatsApp circles, and social media (general-praise-thanks, 13 threads). The most specific praise is for citation grounding: a power user who cross-checks answers against original texts reports "no discrepancies between Qaf's replies and the source hadith/Quran texts" (praise-citation-accuracy, 3 threads). Citation accuracy is the differentiator users notice — which is also why the scripture-misquote and UUID-leak bugs above are disproportionately damaging.
Complaints and critiques. A bilingual user reports the French and Spanish interfaces feel machine-translated and has offered a corrections spreadsheet — a fix workflow is already in motion (translations-feel-machine-translated). Another user criticizes that answers rely on scholars from one particular current, limiting who can benefit; the thread ended in mutual confusion and needs a follow-up (sources-limited-one-school) — this echoes the madhhab-neutrality bug report. Answer length got opposite reactions (too brief vs pleasingly concise); the length setting already exists, so this is a discoverability gap (answer-length-conciseness). A user from Makkah suggests seeking endorsement from bodies like the General Presidency for Scholarly Research and Ifta and offers his personal contacts (seek-institutional-endorsement).
Questions & membership
Lifetime offer is the biggest thread type in the inbox. 25 threads ask for the price, terms, guarantees, and how to buy the $1,000 lifetime plan; several completed Stripe payments, one Belgian organization wants activation on their teacher's account, and a Harvard PhD asked about the deadline (lifetime-membership-inquiries, P0). One purchase was refunded a few hours later under the 90-day guarantee — worth tracking refund rate on the cohort (lifetime-refund-request).
Price resistance is the flip side. Four inquirers declined at $1,000 (one even at the 50%-off $500 for his imam) and say they will wait for the monthly/Pro plan (lifetime-price-too-high). Six more supporters asked for a donation link or small-donation option; two of them balked at the lifetime price and explicitly wanted a cheaper way to give (donation-and-support-requests). That is 10 threads of uncaptured willingness to pay below $1,000.
FAQ candidates — questions people keep asking: - Are the answers religiously sound? Which sources/scholars, which madhhahib, can fatwas be trusted? (7 threads, ai-answer-trust-and-sources; one promoter wants proof against fake fatwas before influencers will promote) - Will Qaf stay free, and how is it funded? (5 threads, free-plan-and-funding-questions) - Which LLMs power Qaf? (2 threads, model-provider-transparency) - Can staff read my conversations? When are chats deleted? (1 thread, unanswered, conversation-privacy-question — sensitive given the personal nature of religious questions) - How do I copy answer text? Is there a voice assistant? Is it on Huawei AppGallery? (copy-answer-text-howto, voice-assistant-question, huawei-appgallery-availability)
Other notable inbound: 10 job/volunteer offers worth keeping as a talent pipeline (inbound-talent-offers); 5 institutional partnership proposals including CILE at HBKU and a Madinah professor (partnership-collab-proposals); 3 offers to evaluate AI quality/safety — directly relevant since Qaf has no eval set (ai-eval-benchmark-collabs); a creator-payout gap blocking payment to a Pakistani Creator Program participant (influencer-creator-marketing); and a security researcher's quran.com/sunnah.com vulnerability reports that need routing (quran-com-security-report-routing).
Recommended next actions
- Fix Arabic-script rendering for Persian/Dari/Pashto/Kurdish (P0). Switch to a font covering the full Arabic script (users suggest Noto Naskh Arabic; one sent a list of standard Pashto/Dari fonts) and fix RTL shaping. 12 reports, an audience relay, and a confirmed two-week-old unfixed report make this the top engineering item.
- Fix the OTP/sign-in pipeline (P0). Reproduce code rejection and resend non-delivery; 9 threads total describe users fully locked out, two of them never answered. Add logging on code validation failures.
- Run an accessibility sprint: label every control for VoiceOver/TalkBack. 16 threads from blind users; a volunteer beta tester is already offering to help. Mostly labeling work, high goodwill return.
- Announce a cheaper monthly tier (or interim donation link) and email everyone who balked at $1,000. 10 threads of unconverted willingness to pay are sitting in the inbox with names attached.
- Publish a public FAQ/trust page covering sources and methodology, madhhab coverage, fatwa disclaimers, free-tier future, funding, model providers, and privacy/data handling. This preempts ~15+ recurring question threads and answers the influencers' "prove it's not fake fatwas" blocker. Answer the unanswered privacy email first.
- Harden answer post-processing and scripture quoting. Strip UUIDs/"DONE"/code artifacts before display, and render Quran/hadith text from the canonical database instead of model output where possible. These defects attack the exact thing users praise most — citation accuracy.
- Ship partial text selection/copy. Top feature request (10 reports) and already promised to several users.
- Take up one of the three eval-collaboration offers (IslamicMMLU, safety evals, LLM observability). Qaf has no eval set; free expert help is on the table and also feeds action 5's trust story.
Methodology
702 Gmail threads were pulled from the shared Qaf inbox using the query {to:qaf.ai cc:qaf.ai deliveredto:qaf.ai} (catch-all forwarding puts every @qaf.ai address in one mailbox), covering 2026-03-23 to 2026-08-24. Each thread was classified by AI agents into categories (bug, feature request, question, membership, feedback, and non-customer categories such as internal system mail and promotions); threads raising multiple issues were counted once per category, yielding 728 classified items. Customer-relevant threads were then clustered by issue, prioritized (P0–P3), and written up as per-cluster detail files with quotes and thread references under feedback/ (index). Much of the mail is in Arabic, with some French, Farsi, and other languages; quotes in this report are translated.