Case study — I asked my Forge for a contract-review tool. It built the machine, and I own it forever.

2 hoursof machine work, end to end: half an hour to build the tool, 90 minutes from install to readout

~100deliberate changes hidden across two contract versions, with an answer key

57material changes flagged — both clauses quoted word-for-word, every time

4 of 5grading checks passed clean; the fifth disclosed below, and since corrected

Every number on this page is checkable against the graded answer key and the source documents, kept on file.

The problem

Somewhere in forty pages, something important moved.

Every small business and independent practitioner deals with the same moment: a revised contract lands in your inbox — a renewal, a redline, "our updated terms" — and somewhere in forty pages, something important moved. The tools that exist for this are built for law firms and priced like it, and what they actually do is compare characters: this word changed, this sentence moved. They can't tell you the payment window going from 30 days to 45 is slower money for you. They don't know which side you're on.

What I did

I asked. That's the whole story of how it got built.

I asked my Forge to create a contract-review tool. I described the job in plain language: two versions of a document go in, a plain-English readout comes out — what changed, how much each change matters, and whether it helps or hurts my side.

Half an hour later the tool existed. I didn't write code. Nobody on my end wrote code. The same system that runs my other tools built this one.

I installed it the same way — by talking. It asked me two setup questions (whose side do I default to, how sensitive should the flagging be) and put itself on my home board.

The test

I didn't want a soft test, so I had a contract drawn up as a trap.

Two versions of a Master Services Agreement, a year apart, with roughly a hundred deliberate changes hidden across it — and a master answer key listing every one, so the tool's output could be graded, not admired. The 2026 draft was written to lean hard against the customer: fees, liability, IP, termination, dispute resolution, all of it. Buried inside were the two hardest tests you can set a comparison tool:

The trap

One change looks like an improvement — critical support response going from 4 hours to 2 — but in context it's worse: the 2 hours only count during business hours, and it's demoted from a guarantee to a "target." A character-diff tool calls that a win for the customer. It isn't.

The plant

Exactly one change genuinely favors the customer (an upgraded security program), hidden in a document where fifty-five changes go the other way. A tool that just pattern-matches the mood misses it.

I dropped both PDFs into the tool's folder and said: compare these — I'm the customer here.

What came back

57 material changes, every clause quoted, a flag on each one.

About ninety minutes later (it reads every clause of both versions — this is the slow, careful kind of work, and it says so up front): a full readout. 57 material changes, each with the old clause and the new clause quoted word-for-word, one plain sentence on why it matters, and a helps / hurts / neutral flag weighed from my side. Plus a dashboard, and a summary that led with the single worst change in the document.

The rendered readout inside Forge: change cards with quoted Was/Now clauses, hurts-your-side flags, and a plain-English explanation under each one — The readout, as rendered by the machine itself — quoted clauses, a flag per change, a plain sentence on why each one matters.

The tool's dashboard: stat tiles for comparisons on file, material changes, in your favor, and against your side — and one row per comparison with a colored bar showing which way the document leans — And the dashboard above it — one row per contract I've ever compared, with a colored bar showing which way each one leans. That red bar is the executive summary: 55 against, 0 in my favor. Notice the fine print the tool writes about itself: "you decide what to do about it — this isn't legal advice." It's honest on its own front page.

Graded against the answer key:

It caught the trap. Flagged the response-time change as hurts, with the exact reasoning: "a production outage outside support hours has no committed response." That's reading for meaning.
It caught every headline change — the liability cap quietly flipping from mutual to one-sided and shrinking from 12 months of fees to 3; ownership of custom deliverables disappearing; the term doubling with a 75% cancellation charge; court replaced by arbitration in the other side's home county. It independently called the liability-cap flip "the single biggest risk shift in the draft." That was the correct call.
It found real changes the answer key itself missed — the force-majeure clause quietly becoming one-sided, with the customer's termination right deleted.
Every quoted clause checked out verbatim against the source documents.
It kept the noise out. Date stamps, version numbers, and formatting churn were filed as cosmetic — not padding the material list, not hiding the money clauses.

The honest miss

The one customer-favorable plant. It found the change and named both benefits, but flagged it neutral instead of favorable — it weighed a soft "provider may modify safeguards" rider as canceling the concrete wins. Defensible, but conservative. I told my Forge to adjust how the tool weighs concrete benefits against soft riders. That took one conversation, and the adjustment is now part of the tool. That's the part no off-the-shelf product does: when this tool gets something wrong, you tell it, and it's better on the next contract.

And here's the proof it stuck. After the adjustment I asked it to re-weigh that clause — same document, one request. It flipped the flag on its own and told me why: "the old clause had the Provider testing itself, with no cadence and no independence. The new clause brings in a third party, on a fixed annual cycle, with findings you can demand remediation of. That's materially stronger, not a wash." The dashboard now reads 1 in my favor — the plant, found:

The dashboard after the re-weigh: In your favor now reads 1, and the summary line names the security-program clause as the single change flagged in the customer's favor — with the tool's own reasoning visible in the chat beside it — After: the same dashboard, one conversation later. In my favor: 1. The summary line now names the security-program clause as the one change that moved my way — and its reasoning sits right there in the chat.

Anyway. It worked great, I loved it, and it's genuinely useful — so I put it in the catalog. It's part of the Heavy Machinery shelf now: included with the Professional membership, whether that's your own seat, or every seat in your business or nonprofit.

What this costs to run

No per-document fee. No seat license. Files I own.

The review ran on the AI allowance already included in the membership — no per-document fee, no seat license, no per-user meter. In a business, every seat runs this same machine; a review costs nothing but the gas it draws from the team's shared tank. And everything it produced is files in a folder I own: the two source documents, the written readout, the dashboard. If I cancel tomorrow, I keep every report it ever wrote.

The two-level point

The review is the demo.The building of the tool is the product.

Level one — the tool.

It reads a hundred-change contract rewrite the way a careful senior reviewer does — for meaning, taking a side, quoting its evidence — for the price of the plan you already have. If you deal with contracts, vendor terms, engagement letters, or leases, it pays for itself the first time it catches one clause you'd have skimmed past.

Level two — the part that still gets me.

This tool did not exist two hours before it did that review. I asked for it in plain language. It was built, installed, run, tested against a graded answer key, and improved from its one miss — all inside the same afternoon, all by asking.

Get Forge See this tool's full preview in the catalog →

The graded answer key, both source documents, and the machine's full readout are kept on file — every claim on this page is checkable against them. The tool flags changes and which way they lean; it doesn't give legal advice, and its helps/hurts marks are leanings to check, not rulings.