ClearTerms

The problem

Financial agreements routinely ask people to accept dense terms under time pressure. A 2017 Deloitte consumer survey found that 91% of respondents accepted legal terms without reading them. The Consumer Financial Protection Bureau maintains a database of agreements from more than 600 card issuers, illustrating how common and varied these documents are. A penalty-rate clause buried several pages deep can raise borrowing costs after one missed payment.

Existing tools do not solve this for the person about to sign. Contract platforms and machine-learning annotation pipelines serve legal professionals; AI summarizers compress agreements into short paragraphs, and compression removes context. When a summary says “there may be fees,” the reader cannot see which clause says so, how much, or under what conditions. The gap was not making documents shorter. It was making them understandable while preserving every word.

Product opinion: annotate, don't translate

The instinctive answer — and many AI tools' default — is to rewrite the agreement in simpler language. I rejected that approach. Financial terms are precision instruments. When a clause says “the fund reserves the right to impose a deferred sales charge of up to 5.5%,” every phrase carries legal weight. “Reserves the right” means the issuer can act but does not have to; “up to” names a ceiling, not a fixed rate. Translating that into “you might pay 5.5% when you sell” strips away conditions that matter.

ClearTerms keeps every word of the original intact and adds a comprehension layer beside it. When users sign, they sign the real agreement — not an AI reinterpretation.

AI should expand human understanding, not replace the source of truth.

Designing AI behavior: prompt engineering as UX

The core design work did not look like a sequence of wireframes. I iterated prompts. Each version changed not what the interface looked like, but what the system did:

The shift from V1 to V3 was an AI-behavior iteration, not a visual one. Each version changed the system's boundaries and priorities. This is prompt engineering treated as interaction design.

Writing the prompt as an interface specification also made failure modes easier to review. I could ask concrete questions of each version: Did it preserve the source? Did it rank clauses selectively? Did it expose uncertainty? Did it cross the line from explanation into advice? The prompt became a testable design artifact rather than an implementation detail hidden behind the screen.

The interface

The layout is a split view: the full original agreement on the left, untouched; annotations on the right. A summary card gives a plain-language overview with impact badges — red directly affects money, yellow is conditional, blue is standard boilerplate — so readers know where to focus before reading every clause. Annotations stay collapsed by default and expand on demand, keeping the interface quiet until detail is needed.

The linking is bidirectional. Selecting a highlighted clause opens its annotation; selecting an annotation scrolls the original clause into view and pulses it. The source and explanation do not lose each other.

Progressive disclosure controls the information density. The collapsed state carries only an impact level and a one-line explanation; the full rationale appears after the reader chooses to inspect it. That hierarchy supports two modes without creating two products: scanning for risk and reading a specific clause closely. The interface does not force every reader through every annotation before they can understand the document's structure.

Where AI must stop

When the model is not confident about a clause, ClearTerms does not hide that uncertainty. Low-confidence annotations use a dashed border and muted text, with a direct warning to consider consulting a financial advisor. In finance, a wrong interpretation can cost real money; saying “I'm not sure” is safer than presenting a guess as fact.

• A Show Original Only toggle removes every AI layer. The source remains the authority.
• Every annotation includes feedback controls, so the reader can contest the model's interpretation.
• The system never advises signing or not signing. That decision belongs to the person reading.

Beyond clause-by-clause reading

Two outputs turn the document into questions a reader can act on. Hidden Costs gathers fees that are scattered across pages into one view. Questions to Askgenerates questions from the agreement's actual clauses, and every question links back to the clause that produced it. The reader can arrive at a bank with specific questions instead of a generic summary.

Traceability remains the rule in both views. A hidden cost is useful only when a reader can return to the language that creates it; a suggested question is useful only when the bank employee and customer can inspect the same clause. ClearTerms therefore treats summaries as navigation into evidence, not as replacements for it.

Limits, and what's next

ClearTerms currently handles English only and works best with North American agreements. The AI-behavior design went through structured iteration, but the product has not yet been through a formal usability evaluation. That evaluation is the next research step. The product uses React, Tailwind CSS, and the Claude API and is deployed on Vercel — designed, built, and shipped solo.