I once asked an AI to build a brand website. What came back looked convincing at first glance, but it was broken in the ways that actually matter. The header showed exactly two links. There was no "About" page anywhere. An admin panel existed on the backend, yet no button or route on the frontend could actually reach it. The server side functioned. The user side did not.

Most people who run into this assume the AI got lazy or hit a token limit. That is not what is happening. The issue is structural. AI coding tools are built to police internal consistency. They ask: "Does everything I declared line up with itself?" They do not ask: "Does this kind of deliverable contain everything it is supposed to contain?" If you mention two pages for your brand site, the model dutifully checks whether those two pages link back to each other. Once the links resolve, it considers the task complete. It has no inherent notion that a brand site needs an About page, trust signals, or a contact path. There is no standard in its head.

I call the fix a Completeness Baseline.

A completeness baseline is simply a standard checklist of what a given deliverable must contain. You generate the artifact, then you check the real output against this external standard instead of trusting the tool's internal sense of "done."

Checking Across Three Layers

Not all missing pieces are equally obvious. A useful baseline checks three distinct layers.

Existence. Does the part actually exist? This sounds basic, but an AI will happily build a navigation wrapper that references a dashboard page without ever generating the dashboard page itself. The reference exists. The target does not.

Reachability. Can a real user actually get to it? Hidden admin panels are the classic symptom. The route and the component might both exist in the codebase, yet no menu item, button, or redirect exposes them to the interface. If a user cannot trip over it during normal use, it is not really there.

Substantiation. Is there real data or structure behind the surface? A page that loads but contains only placeholder text and no image upload slot is a skeleton wearing a costume. An About page without an image slot or editable text field is not complete, even if the HTML renders cleanly.

These three layers catch different classes of gaps. Existence finds the missing shoe. Reachability finds the shoe locked in a closet. Substantiation finds the shoe with no sole.

Baselines by Type

One baseline will not cover every project. You need to classify what you are building and define the non-negotiables for that category.

Brand sites need specific sections and reachable pages. Think About, Contact, privacy links, and navigation that actually exposes every major view.

APIs need documentation, exhaustive error codes, and rate limits. A working endpoint that returns 200 OK for success and a generic 500 for everything else is not a finished API. It is a hazard.

Automations need logs and failure alerts. If a workflow breaks at 2 AM and nobody knows until a human checks the run history, the automation is incomplete. Observability is not a bonus feature. It is part of the deliverable.

Once you name the category, the baseline writes itself. The hard part is enforcing it after the code is generated.

Design Rules That Hold Up

I rebuilt my own workflow around this idea and landed on three practical rules that keep projects from slowly unraveling.

Capability-gated design. Before you commit to a tool or a generated module, detect what it can actually do. If a component library lacks native mobile drawer support, do not let the AI generate a full navigation scheme that assumes it exists. The system should degrade gracefully instead of breaking when a capability is missing. Know the boundaries first. Design inside them.

Public engine, private values. Use a public engine for the heavy lifting, but inject your private data, configs, and conventions at runtime. This keeps your personal or company methods safe and separate from the generated scaffold. The AI builds the frame. You snap in the glass. This prevents the model from hard-coding assumptions that violate your actual standards or leak sensitive patterns into public training contexts.

Ajukan usulan, jangan lanjut otomatis. Biarkan AI menyarankan langkah berikutnya, tetapi paksa manusia untuk membuat pilihan. Otomatiskan alur eksekusi, jangan pernah penilaiannya. Ketika sebuah model secara otomatis menghasilkan migrasi database, skema autentikasi, dan hook pembayaran dalam satu tarikan napas, Anda mendapatkan kenyamanan dengan mengorbankan pengawasan. Buatlah alat tersebut menyajikan rencana. Biarkan pengembang yang menekan tombolnya.

Memilih Model yang Tepat untuk Pekerjaan Tersebut

Saya menguji hal ini di berbagai model berbeda dan menemukan perbedaan nilai yang jelas. Model yang lebih murah menangani tugas-tugas mekanis yang berat dengan sangat baik. Mereka menghasilkan boilerplate, komponen berulang, dan stub struktural lebih cepat daripada Anda mengetik. Model yang mahal baru sepadan dengan harganya ketika mereka menunjukkan pengendalian diri. Anda menginginkan opsi premium saat ia menolak untuk mengarang perbaikan palsu dan justru menandai adanya celah yang nyata. Model yang berhalusinasi menciptakan solusi sementara untuk endpoint API yang hilang sangatlah berbahaya. Model yang berhenti dan berkata, "Alur kerja ini memerlukan target webhook yang belum ditentukan," barulah layak dibayar. Bayarlah untuk ketajaman penilaian, bukan untuk volume.

Bangun Sistem, Jangan Mengejar Keajaiban

Jangan mencoba memperbaiki celah AI dengan menggunakan model yang lebih besar atau lebih banyak trik prompting. Perbaikilah dengan sebuah baseline. Celah tersebut bukanlah masalah kapasitas. Itu adalah masalah ekspektasi.

Untuk menghentikan hilangnya blok kode, lakukan tiga hal.

Berikan sistem sebuah baseline kelengkapan sebelum kode apa pun ditulis. Jadikan itu daftar periksa (checklist) fisik yang berdampingan dengan tugas tersebut.

Tanamkan konvensi Anda ke dalam bagian-bagian yang dapat digunakan kembali. Jika Anda menerapkan baseline melalui template, aturan lint, atau scaffold yang sudah jadi, AI akan memulai dari posisi yang benar alih-alih hanya berharap ia bisa menebak standar Anda.

Otomatiskan alur kerja sambil tetap membiarkan manusia memegang kendali penilaian. Biarkan mesin menangani pengulangan. Simpan pengambilan keputusan untuk orang-orang yang memahami konteksnya.

Berikut adalah satu kebiasaan terakhir yang mengubah cara saya bekerja. Jika Anda membuat keputusan desain yang sama sebanyak tujuh kali, berhentilah menganggapnya sebagai pilihan sekali jalan. Itu bukan pengulangan. Itu adalah sebuah hukum. Beri nama. Ubah menjadi aturan. Dokumentasikan. Saat Anda mengodifikasi pola tersebut, Anda menghilangkan kemungkinan AI akan menyimpang darinya pada kali kedelapan.

Sumber untuk kerangka kerja ini dan eksplorasi aslinya dapat ditemukan di sini.

Jika Anda ingin bertukar pikiran dengan orang lain yang menghadapi masalah yang sama, Anda dapat bergabung dengan komunitas belajar GyaanSetu.