AI Wrote 80% in 10 Minutes. The Last 20% Took 6 Hours.

AI can build the happy path in minutes. It cannot build the reality of production.

I tracked 47 features using an AI agent. I measured two things:

  • Generation time: From the first prompt to a finished PR.
  • Ship time: From the PR to the actual merge.

The split is almost always 80/20. The AI does 80% of the work fast. The last 20% takes the most time.

The AI writes code for the prompt you give it. The delay comes from what you did not mention. It comes from things you forgot to think about.

The missing 20% usually falls into five categories:

  • Empty states: What does the UI look like when a user has no data?
  • Error handling: What happens when the network fails or an API returns a 500 error?
  • Domain edge cases: Specific rules like local payment formats or legacy data.
  • Performance: Code that works for 50 rows but breaks at 5 million rows.
  • Maintainability: Code that works today but is hard to change tomorrow.

I stopped treating AI as a magic wand. I started using these four rules to save time:

  1. Budget 4x time. If the AI says a task takes 10 minutes, plan for 40.
  2. Prompt for the unhappy path first. Ask the AI to handle empty inputs or network failures before it writes the main logic.
  3. Write failure tests first. Define what should break before the AI generates the code.
  4. Keep a 20% journal. Write down why the last part of a feature took so long. This helps you spot patterns.

AI makes your typing faster. It does not make your shipping faster unless you think ahead. The goal is not to write prompts quickly. The goal is to move features into production.

Think about the problem before you open the AI tool. If you cannot answer what a user might do wrong, do not start coding yet.

Source: https://dev.to/susiloharjo/ai-wrote-80-in-10-minutes-the-last-20-took-6-hours-5764

Optional learning community: https://t.me/GyaanSetuAi