Vibe Coding a Harvey Competitor: Fun Demo, Expensive Reality

Summary of an essay from Dorna Moini's Legal Tech Notes, published on Substack on January 27, 2026. Dorna Moini is the CEO and Founder of Gavel and a former litigator at Sidley Austin. This page gives the argument and the numbers; the full essay is on Substack.

In brief: Viral posts about someone who "built a Harvey competitor" in an afternoon skip the hard part. Generating text is easy; making AI work reliably inside Microsoft Word on a 90-page contract with tracked changes, defined terms and exhibits is the actual product. The essay gives five reasons weekend-built tools stop at the demo, and a clear rule for when a firm should build its own anyway: narrow, internal, bounded problems with nothing off the shelf that fits.

Why is "inside Microsoft Word" the real test?

Lawyers work in Word, and most viral tools run in a chat box. A .docx file is packaged Office Open XML, and edits that touch structure, numbering, cross-references, tables, tracked changes, comments or defined terms are fragile. Word's own APIs use an asynchronous, batch-based model, so the order of edits matters, and an agent has to stay in sync with the live document on the lawyer's laptop, securely. The essay cites Jamie Tso, the vibe coder interviewed by Richard Tromans of Artificial Lawyer, on hitting exactly these limits with no-code tools.

Why does the demo break on real contracts?

Demos use short, clean documents. Real contracts are long and layered with defined terms, exhibits, tables, numbering quirks and tracked changes. A production system has to handle that length without slowing or crashing, manage context and chunking so the model stays accurate, and preserve formatting on the way back. Tightly scoped tools work; generalizing across document types, practice groups or jurisdictions raises complexity faster than linearly.

Training data is the next wall. Market-standard redlines, playbook positions and fallback clauses do not emerge from a general model. At Gavel, the team worked with practicing attorneys to build dozens of playbooks from real deals and gold-standard templates; without that grounding, output sounds plausible and misses how lawyers actually negotiate.

What are the security risks of a vibe-coded legal tool?

Legal documents are among the most sensitive data a firm holds, and vibe-coded tools rely on prompts and untrusted inputs. Better infrastructure fixes the obvious failures, such as leaked credentials and sloppy file handling, but prompt injection is structural: a contract or upload is text the model reasons over, so instructions hidden in a document can steer the model. The UK National Cyber Security Centre warns it is unlike SQL injection and cannot be patched with a clever prompt, and OWASP ranks it the top risk category for applications built on large language models. Studies of user-built GPTs have found instruction leakage and unintended exposure of attached files.

What is the hidden cost after launch?

Maintenance. Models and providers change often and new failure modes appear, while established companies ship improvements weekly, so the gap widens. Tso himself notes that once someone forks a repo and builds on it, maintenance is theirs. Reliability comes from evaluation frameworks, test sets, regression testing and feedback loops, plus an interface that lets a lawyer audit each change, see its source, and accept or reject it. At Gavel, agents working in real time in Word were hitting Office's performance limits, so for Gavel Exec the team built its own APIs on top of the OOXML standard rather than relying on Microsoft's.

When should a firm build its own tool anyway?

Build when the problem is narrow, internal and clearly bounded, and nothing off the shelf does what you need. Buy or partner when the workflow is used constantly, is mission-critical or client-facing, lives in Word, touches sensitive data, or has to be dependable across many matters and users. The essay's closing line is that the valuable part of legal software is the part that is boring, dependable and safe.

Key numbers and sources from the essay

• Prompt injection: ranked the top risk category for LLM applications by OWASP; the UK National Cyber Security Centre says it cannot be patched with a prompt.
• Gavel built dozens of contract playbooks with practicing attorneys from real deals and templates.
• Gavel Exec runs on Gavel's own APIs built on the Office Open XML standard, because Office's native APIs were designed for human-speed input.
• Source interview: Richard Tromans (Artificial Lawyer) with Jamie Tso on the limits of vibe-coded legal tools.

What this means for a law firm or legal team

1. Keep a list of the narrow internal problems where a weekend tool is the right answer: a matter-intake checklist, a clause finder for one template set, an internal FAQ bot.
2. For anything client-facing or in Word, evaluate vendors on the boring questions: document length limits, tracked-changes handling, prompt injection defenses, and how often they ship.
3. Ask any vendor what their test set looks like and how they catch regressions when a model provider changes.
4. Gavel Exec is the Word-native version of what the essay describes; the security page covers data handling, and the contract review comparison shows which competing tools work inside Word.

Frequently asked questions

What is vibe coding?

Building software by describing what you want to an AI coding tool and accepting what it generates, with little or no hand-written code. It produces working demos quickly and struggles with reliability, security and maintenance.

Can a law firm build its own AI contract review tool?

For a narrow internal use, yes. For client-facing work inside Microsoft Word on long contracts with sensitive data, the essay argues the integration, training data, security and maintenance costs make buying or partnering the better choice.

What is prompt injection and why does it matter for legal AI?

Prompt injection is when instructions hidden in a document or input steer the model's behavior. Contracts and uploads are text the model reasons over, so a tool that reads them is exposed. OWASP ranks it the top risk for LLM applications.

Read the full essay: Vibe Coding a Harvey Competitor: Fun Demo, Expensive Reality on Substack, or subscribe to Dorna Moini's Legal Tech Notes.

Related on gavel.io: Gavel Exec, Gavel security, Best AI contract review tools.