At 11 p.m., I nearly mistook “it runs” for “it’s ready to ship”

Last Wednesday at 11 p.m., I was updating the quote page for a freelance client in Hangzhou. AI built the page quickly, and it looked convincing. I sent it the next day, only for the client to point out that a key button wasn’t recording the source. The awkward part: if I keep making mistakes like that, I’ll lose trust before I gain efficiency. I’ve been stuck here too. I kept asking, “Did I build it?” instead of, “Can I prove it works?” If we take on websites, small tools, or automation projects, this check is worth adding.

What matters more now is writing clear requirements and acceptance checks

In a recent Pragmatic Engineer interview, Microsoft CEO Satya Nadella put it bluntly: as code generation gets cheaper, verifying that the code meets the requirements becomes the harder part. That shift also affects nontechnical founders. Lin Xiao, who still uses Linear, runs a three-person design studio in Hangzhou. She has AI write the functional specification first, then list steps a person can check manually. Once development is done, she confirms each item on the checklist instead of judging whether the page “looks right.” She doesn’t know how to code and didn’t buy an expensive system. She simply shifted the time she used to spend verbally confirming requirements to the period before delivery.

Replication cost: $0, about 90 minutes, no coding required

Money: $0, using an AI chat tool we already have and a regular document. Time: About 90 minutes across three sessions: 20 minutes to define the goal and boundaries, 30 minutes to list acceptance steps, and 40 minutes to have AI compare the build against them and make revisions. Technical barrier: We need to describe the business outcome, not understand code. First step: Open the AI tool we already use, create a document, and enter: “What am I helping the client accomplish, what counts as success, and what absolutely must not happen?” Then ask AI to turn it into five to ten acceptance checks a person can verify, and tick them off one by one. We can also attach the client-approved requirements to the checklist so both sides aren’t relying on half-remembered agreements. There’s no urgency to try this immediately—manual checks still work. This method simply reduces rework.

Here’s how I’d set it up at each stage

Just starting out: If I’m building a side project or taking on my first few jobs, I’d begin with a handwritten acceptance checklist in a spreadsheet rather than building a complicated process. For each project, I’d pick the three outcomes that most directly affect getting paid.

With one or two clients: If the same requirements keep coming up, I’d have AI generate a standard checklist from past conversations, then ask the client to confirm it before work begins. That turns “I thought we agreed” into something both sides have seen.

Scaling up: If I’m part of a one-to-five-person team, I’d keep the requirements, acceptance checks, and revision history in one project document, with one person responsible for final sign-off. The goal isn’t faster output for its own sake. It’s making sure the team knows what to trust and what to check.

As code generation gets faster, what we really need to protect is whether the client trusts us to stand behind the result.