AI writes exactly what you say — but not what you need

You ask AI to build a feature. It asks exactly two questions, then dives into code. Two days later, you receive software that runs beautifully — matching your words to the letter, yet completely wrong for what you actually need. This is not the model's fault. The fault lies in nobody asking enough questions before the work began. This article explains why the requirements spec is becoming the most important skill of the delegation-to-AI era.

A situation every engineering team using AI has hit

The scenario plays out increasingly often at Vietnamese companies: a developer opens an AI tool and types something like "add a report export feature for customers." The AI responds briskly with exactly two clarifying questions, gets a few lines of answers, and starts coding.

The result after two days: the feature works. Clean code, proper syntax, even tests included. But when the business unit tries it out, everyone is stunned: a critical data field is missing, the workflow doesn't match real operations, and nobody accounted for the edge cases staff deal with every day.

The AI built exactly what you said — but that wasn't what you needed. The gap between "said" and "needed" is the dead zone of every software project, and AI now exposes that gap faster than ever.

The AI didn't do it wrong — it did a wrong requirement right

The first thing to acknowledge: this is not the model's fault. AI works on the principle of receiving instructions and executing them. If the instruction is a single sentence plus two clarifying answers, it will fill in the rest of the requirement with the assumptions that seem most reasonable to it — not with your company's business reality.

An experienced engineer, given a vague requirement, will reflexively fire off a barrage of questions: where does the data come from? Who can view it? Does history need to be kept? What happens when a customer has no transactions? AI doesn't do that on its own — it is designed to follow instructions, not to "interrupt" and probe like a true consultant.

Put differently: the faster you delegate, the more carefully you must prepare. AI's coding ability far outruns its mind-reading ability.

What a spec is, and why it's back in the starring role

A spec (short for specification — a requirements document) is a written statement of what a feature must do, for whom, under what conditions, and what counts as done. Before the AI era, large development teams always wrote specs — but many small teams treated them as pointless paperwork, preferring to just "tell each other verbally" for speed.

AI fundamentally changes that equation. Humans hear requirements through accumulated business experience; AI hears requirements through exactly what is written down. When the only source of truth between you and the AI is text, the quality of that text determines the quality of the product. A good spec doesn't make AI code faster — it helps AI get it right the first time, and that is what truly saves time, because fixing a feature built on the wrong foundation costs far more than rewriting from scratch.

Three must-haves before delegating work to AI

1. A clear requirements checklist

Before typing a prompt, finish answering the foundational questions:

• What problem does this feature solve, and for whom? (are the end users internal staff, customers, or other systems)
• How does the main workflow run? Describe each step in order, from data in to results out
• What are the edge and boundary cases? Empty data, negative numbers, unauthorized users, slow networks — these are the branches that make a feature "matching everything you said" yet useless
• Scope: what's in and what's out for this version

2. A Definition of Done

This is a familiar concept in professional software development: making explicit the criteria for considering a feature finished. With AI, write it as a verifiable checklist — for example: "when the user clicks export report, the downloaded file contains exactly columns X, Y, and Z; processing time doesn't make users wait too long; a friendly message shows when there's no data." With a Definition of Done, you and the AI aim at the same target, and reviewing the output becomes objective rather than a matter of taste.

3. Technical and non-technical constraints

AI has no idea your business has its own limits unless you say so:

• Technology: language, framework, and version must follow the existing system
• Information security: no logging passwords, no exposing customer data, comply with internal policies
• Integration: must be compatible with existing APIs, databases, and naming conventions
• Performance and resources: which servers it runs on, what load it must handle

Turn it into a process, not "prompt and pray"

The encouraging sign is that the engineering community is tackling this systematically. Matt Pocock — author of the well-known Total TypeScript course — recently pushed a GitHub repo dedicated to the problem of "AI coding the wrong thing because the requirement was unclear," aimed at turning delegation to AI into a process with serious document preparation, rather than a game of chance.

This trend deserves attention from Vietnamese businesses: the competitive edge in the AI era isn't about which team has the more powerful tools (everyone can use those) — it's about which team prepares better inputs and controls output quality more tightly.

A perspective for Vietnamese businesses

• Train your team to write specs. It's a one-time investment that pays off for delegating to AI, to new hires, to outsourcing, or to software partners alike.
• Apply the principle "treat internal projects like outsourced projects." If you couldn't hand a task to another company in a single sentence, don't hand it to AI in a single sentence either.
• Keep tight control over major changes. AI produces code very fast, so the human approval boundary needs to be clearer, not looser.
• Invest in the operations layer, not just the prompt layer. Database account permissions, environment separation, an approval gate for operations, audit logs — this is a business's real fence.

Conclusion

AI is the best executor software engineering has ever had — but it is only as good as the requirements it receives. The story of "AI asks two questions and codes the wrong thing" will keep repeating at every company that treats delegation to AI as a shortcut skipping the preparation step. Reverse the order: write a clear spec first, delegate to AI second. That is the smallest investment that turns AI into a genuine advantage instead of a code factory whose output needs redoing.

Source: compiled from technical discussions on Facebook, edited by Wotech Electronics.

Is your business preparing to bring AI into its software development process? Contact Wotech for advice on setting up the right workflow, requirements documentation standards, and technical infrastructure.