We Tested a Generic AI Against HAQQ on Real Startup Legal Documents. Here's What Happened.

A real experiment with a real client: Claude (generic LLM) vs HAQQ (domain-specific legal AI) drafting co-founder agreements and IP assignments. 5 rounds, 13 issues found, 32 pages produced. The results weren't even close.

The Setup

A founder came to us with a concrete need. They're co-founding an AI startup with a technical partner who built the entire codebase. The business founder handles strategy, growth, and partnerships. They needed two documents before they could move forward:

• IP Assignment Agreement — Who owns what when the company incorporates?
• Co-Founder Agreement — Equity split, vesting, decision rights, the works.

These aren't hypothetical. The founders are actively pursuing multiple paths: an acquisition listing, open-source launch, and a pre-seed raise. The documents need to be real.

We proposed an experiment: What if we had a generic LLM draft the documents first, then had HAQQ review and revise them? The founder agreed. Here's what happened across 5 rounds.

Round 1: Claude (Generic LLM)

The founder gave Claude Opus detailed context about the startup — the tech stack, the equity split, their open-source commitment, German jurisdiction, the three strategic paths — and asked it to draft both agreements.

What came back: A ~10-page document covering the basics. IP assignment, equity split, vesting schedule, open-source clause, German arbitration. It looked like a legal document. It used legal language. It had section numbers.

A solid first draft from a smart intern who's read a few term sheets.

For a generic AI with no legal training data, no jurisdiction-specific knowledge, and no understanding of startup mechanics, it was impressive. Two years ago, getting this output from any AI would have been headline news.

But the founder wasn't looking for impressive. They were looking for signable.

Round 2: HAQQ Reviews and Revises

We took Claude's draft and fed it into chat.haqq.ai.

HAQQ's first response wasn't a revised document. It was a 13-point critique.

What HAQQ Found Wrong with the Generic AI Draft

• No pre-incorporation mechanics — Who holds the IP before the company exists?
• Absolute warranties instead of knowledge-qualified reps — 'I represent that I own all IP' vs 'To the best of my knowledge' — the first is a lawsuit waiting to happen
• No third-party/OSS carve-outs — The codebase uses open-source libraries. The draft would have assigned their licenses to the company.
• Missing Good Leaver / Bad Leaver distinction — Standard in European VC
• No vesting suspension clause — What if a founder takes a 6-month break?
• Wrong license default for patent-sensitive code — MIT was used. For AI with potential patents, Apache 2.0 is the right choice (explicit patent grant).
• No ROFR mechanics — No right of first refusal structure at all.
• No FMV determination process — How do you determine fair market value in a buyout? Silence.
• Jumps straight to arbitration — No graduated dispute resolution.
• No drag-along / tag-along provisions — Critical for an acquisition marketplace sale path.
• No side projects policy — Both founders have other ventures. No carve-out language.
• No dissolution IP license-back — If the company shuts down, can founders use their contributed IP?
• No confirmatory patent schedule — No mechanism for future patent filings.

Every single one of these is the kind of thing a startup lawyer would catch in the first read. None of them are exotic. They're table stakes for a real founder agreement.

The Revised Version

HAQQ then produced a 27-page revision across two interlocking agreements with schedules.

Agreement 1: IP Assignment (14 sections + 2 schedules)

• Pre-incorporation trust mechanics (Section 3.2)
• Third-party and OSS carve-outs (Section 4.2)
• Apache 2.0 as default for patent-sensitive code (Section 6.2)
• Knowledge-qualified representations (Section 11.1)
• Open-source governance framework (Section 8)
• Founder license-back on dissolution (Section 9)
• Schedule 1: Detailed founder-contributed IP (7 technical categories, 6 strategic categories)
• Schedule 2: Patent matters and confirmatory steps

Agreement 2: Co-Founder Agreement (17 sections)

• Equity split with 4-year vesting, 1-year cliff
• Good Leaver / Bad Leaver with different buyback pricing
• ROFR: Company-first, then non-selling founder secondary
• FMV via independent valuer under German Arbitration Institute (DIS) rules
• Broad-based weighted-average anti-dilution (not full ratchet)
• Pro rata participation rights
• Narrow non-compete with explicit OSS contribution carve-out
• 4-step deadlock resolution (internal → advisor → mediation → binding arbitration)
• Dissolution with open-source releases surviving
• Perpetual license-back for each founder's contributed IP

Round 3: Claude Reviews HAQQ's Work

We then brought HAQQ's 27-page revision back to Claude for a neutral review. Would a generic AI recognize the improvements? Or would it think its own draft was fine?

Claude's assessment was unambiguous:

HAQQ's revision is a major upgrade. It transforms what was a reasonable AI-generated template into something approaching sign-ready. Grade: B+ to A-.

Credit where it's due — Claude was honest. It identified 7 remaining refinements:

• Quantify vesting suspension triggers (define 'material break' numerically)
• Add a deemed resignation clause for founders who drift away without formally leaving
• Shorten deadlock timelines (30 days is too slow for a 2-person startup)
• Add a named carve-out for the founder's other venture in the side projects policy
• Add the founder's dashboard contribution to Schedule 1
• Consider a pre-financing drag-along for acquisition-marketplace-style exits
• Complete the technical founder's full legal name

Good suggestions. But here's where it got interesting.

Round 4: HAQQ Reviews Claude's Review

We fed Claude's 7-point feedback back into HAQQ. Would HAQQ agree? Push back? Find things Claude missed?

HAQQ's response: "Claude's feedback is 85-90% aligned with what we would recommend."

HAQQ agreed with 6 of 7 points. On compensation, HAQQ partly disagreed — recommending a separate side letter rather than baking salary and sale economics into the co-founder agreement. Smart separation of concerns.

But then HAQQ went further. It flagged 5 Germany-specific issues Claude completely missed:

• GmbH share transfer mechanics — German law often requires notarization for share transfers
• "Held in trust" doesn't map to German law — Anglo common-law trust concepts don't translate — needs treuhänderisch language
• Moral rights can't be broadly waived under German law — German Urheberrecht limits how broadly you can waive moral rights
• Patent inventor chain needs tightening — Confirmatory assignments needed for collaborators
• Tax implications of pre-company IP contribution — Contributing IP into a German entity may have valuation and tax consequences

This is the moment the experiment got real. Claude's feedback was solid in the abstract. But it was reviewing as if this were a Delaware startup with common-law mechanics. HAQQ knew the startup is incorporating in Germany and applied the right legal framework.

Round 5: HAQQ Implements Everything

This is where HAQQ proved it's not just a critic. We fed back Claude's 7 suggestions plus HAQQ's own 5 Germany-specific findings, and asked HAQQ to produce Version 2.

What came back: 32 pages. Two agreements. Three schedules.

HAQQ didn't just apply the feedback. It went beyond it.

Things nobody asked for (but HAQQ added anyway)

• Pre-financing tag-along (Section 6.6) — if one founder sells >20% to a third party, the other has a tag-along right. Neither Claude nor anyone on the team suggested this.
• Pre-incorporation economics (Section 9) — no salary before funding, expense reimbursement framework, founder loan tracking.
• Founder Conduct and Values (Section 10) — shared belief in open-source AI, ethical development, mutual respect. Not legally binding, but sets tone for the relationship.
• Formal role definitions — technical founder as CTO, business founder as CEO, with day-to-day autonomy in each domain.
• Confirmatory schedules with identifiers (Section 8.4) — mechanism to list repos, domains, trademark registrations, and patent application numbers post-incorporation.

The Evolution: 10 Pages to 32 Pages in 5 Rounds

• Round 1 — Claude (generic LLM): Initial draft — decent template (~10 pages)
• Round 2 — HAQQ (legal AI): 13-point critique + full revision (27 pages)
• Round 3 — Claude: Review — 'B+ to A-.' 7 refinements. (feedback)
• Round 4 — HAQQ: Meta-review — '85-90% aligned.' +5 Germany-specific. (feedback)
• Round 5 — HAQQ: Version 2 — all feedback implemented + extras (32 pages)

From a basic template to a 32-page, 3-schedule, German-law-aware, open-source-friendly, multi-path-exit-ready founder document package. Built entirely by AI. Argued by two different AI systems with different strengths.

Why This Matters

This experiment revealed something important: the best results come from making AIs argue with each other.

Claude brought breadth — wide knowledge, honest self-assessment, good structural suggestions. HAQQ brought depth — jurisdiction-specific precision, document architecture, German-law compliance.

Neither alone produced the ideal document. Together, through 5 rounds of iterative review, they produced something a German startup lawyer can finalize in a few hours.

The gap comes from domain knowledge:

• Jurisdiction awareness: HAQQ knows Verfügungsgeschäft from Verpflichtungsgeschäft. Claude treated it like a Delaware C-Corp.
• Document architecture: HAQQ builds interlocking agreements with cross-referenced schedules. Claude treated them as independent files.
• Startup mechanics: Good Leaver/Bad Leaver, ROFR cascades, anti-dilution formulas, vesting suspension — these are patterns HAQQ has seen thousands of times.
• Iterative depth: When reviewing feedback, HAQQ didn't just apply changes — it added entire sections that neither Claude nor the founder thought to request.

The Bottom Line

After 5 rounds of a generic LLM and a legal AI arguing over real founder documents, our client has a 32-page package that's been stress-tested from both sides. Claude brought the breadth. HAQQ brought the depth. Each one caught things the other missed.

That's why we built HAQQ. Not to replace general-purpose AI, but to be the specialist that makes the generalist's output actually signable.

This experiment was conducted in April 2026 using Claude Opus 4.6 and chat.haqq.ai. The client's identity and startup details have been anonymized. No lawyers were harmed — but one will be hired to finalize the documents.

الإعداد

جاءنا مؤسس باحتياج محدد. يؤسس شركة ذكاء اصطناعي ناشئة مع شريك تقني بنى قاعدة الكود بالكامل. المؤسس التجاري يتولى الاستراتيجية والنمو والشراكات. احتاجوا وثيقتين قبل المضي قدماً:

• اتفاقية تنازل عن الملكية الفكرية — من يملك ماذا عند تأسيس الشركة؟
• اتفاقية المؤسسين المشتركين — توزيع الأسهم، الاستحقاق، حقوق القرار، وكل شيء.

اقترحنا تجربة: ماذا لو صاغ ذكاء اصطناعي عام الوثائق أولاً، ثم راجعها HAQQ وعدّلها؟ وافق المؤسس. إليك ما حدث عبر 5 جولات.

الجولة الأولى: Claude (ذكاء اصطناعي عام)

أعطى المؤسس لـ Claude Opus سياقاً مفصلاً عن الشركة الناشئة وطلب صياغة الاتفاقيتين. ما عاد: وثيقة من ~10 صفحات تغطي الأساسيات.

مسودة أولى جيدة من متدرب ذكي قرأ بعض أوراق الشروط.

لكن المؤسس لم يكن يبحث عن شيء مثير للإعجاب. كان يبحث عن شيء قابل للتوقيع.

الجولة الثانية: HAQQ يراجع ويعدّل

أخذنا مسودة Claude وأدخلناها في chat.haqq.ai. استجابة HAQQ الأولى لم تكن وثيقة معدّلة. كانت نقداً من 13 نقطة.

كل واحدة من هذه المشكلات هي من النوع الذي يكتشفه محامي شركات ناشئة في القراءة الأولى. ليست غريبة. هي الحد الأدنى لاتفاقية مؤسسين حقيقية.

ثم أنتج HAQQ مراجعة من 27 صفحة عبر اتفاقيتين متشابكتين مع جداول.

الجولة الثالثة: Claude يراجع عمل HAQQ

أعدنا مراجعة HAQQ المكونة من 27 صفحة إلى Claude لمراجعة محايدة.

مراجعة HAQQ ترقية كبيرة. تحوّل قالباً معقولاً من الذكاء الاصطناعي إلى شيء يقترب من الجاهزية للتوقيع. التقييم: B+ إلى A-.

حدد Claude 7 تحسينات متبقية. اقتراحات جيدة. لكن هنا أصبح الأمر مثيراً.

الجولة الرابعة: HAQQ يراجع مراجعة Claude

أدخلنا ملاحظات Claude السبع في HAQQ. وافق HAQQ على 6 من 7 نقاط. لكن ذهب أبعد وأشار إلى 5 مشكلات خاصة بالقانون الألماني لم يلاحظها Claude إطلاقاً:

• آليات نقل أسهم GmbH — القانون الألماني يتطلب غالباً التوثيق الرسمي
• "محتفظ به في أمانة" لا ينطبق على القانون الألماني — يحتاج صياغة treuhänderisch
• الحقوق المعنوية لا يمكن التنازل عنها بشكل واسع بموجب القانون الألماني
• سلسلة المخترع للبراءات تحتاج تشديداً
• الآثار الضريبية للمساهمة بالملكية الفكرية قبل التأسيس

الجولة الخامسة: HAQQ ينفذ كل شيء

هنا أثبت HAQQ أنه ليس مجرد ناقد. أدخلنا اقتراحات Claude السبع بالإضافة إلى نتائج HAQQ الخمس الخاصة بألمانيا. ما عاد: 32 صفحة. اتفاقيتان. ثلاثة جداول.

التطور: من 10 صفحات إلى 32 صفحة في 5 جولات

من قالب أساسي إلى حزمة من 32 صفحة، 3 جداول، متوافقة مع القانون الألماني، صديقة للمصادر المفتوحة، جاهزة لمسارات خروج متعددة. بُنيت بالكامل بالذكاء الاصطناعي. تناقش فيها نظامان مختلفان بنقاط قوة مختلفة.

لماذا هذا مهم

كشفت هذه التجربة شيئاً مهماً: أفضل النتائج تأتي من جعل أنظمة الذكاء الاصطناعي تتجادل مع بعضها البعض.

Claude جلب الاتساع. HAQQ جلب العمق. كلاهما لم ينتج الوثيقة المثالية وحده. معاً، عبر 5 جولات من المراجعة التكرارية، أنتجوا شيئاً يمكن لمحامي شركات ناشئة ألماني إنهاءه في ساعات قليلة.