When AI can help write code, how can someone who has built a small tool know it is usable, rather than merely appearing to run?

I want to move attention from lines produced to requirements, acceptance, and maintenance. Trying things remains valuable. A project does not need to be converted into an entire engineering team’s workload to explain what was learned.

1. Begin with a problem you can explain

“Build a complete platform” can make discussion vague. I would first ask who has which problem and what action they want to complete. For a hypothetical tool organizing public articles, finding an article and displaying the correct title and source already provides a checkable goal.

This is a hypothetical small tool, not a new product accepted in this essay. It turns requirements into more than a feature list and creates an opportunity to notice mistakes or unhelpful work before expanding.

AI can propose code and explanations, and users can learn about work through specific problems. Describing a requirement and independently evaluating architecture, security, or long-term reliability are different capabilities. They need not be combined into one declaration of competence.

2. Lines of code cannot stand directly for effective delivery

A file may contain repetition, generated data, tests, configuration, or attempts later deleted. Even correctly recorded line counts need a defined counting scope. They cannot directly stand for quality, functional value, or team person-days.

Outsourcing quotes, tool fees, the author’s time, and maintenance costs also have different scopes. Dividing a complete-project quote by only some subscriptions or API expenses does not establish overall cost efficiency.

The earlier article’s days, lines, equivalent person-days, and multiples lack complete records verifiable in this revision, so they are no longer presented as case outcomes. The retained questions are what resources were invested, which acceptance checks were completed, and what remains undone.

3. Divide “it runs” into several checks

A screen opening answers only part of the question. Data correctness, unusual inputs, usability across devices, and detectable failures need checks tied to actual use.

If the hypothetical article tool lists titles, one could also check whether links identify the right articles, updates preserve original content, and missing data is misleadingly displayed as blank. These are possible acceptance scenarios, not software tests performed in this revision.

Private data, permissions, or external actions also call for appropriate professional checks and explicit authority. Noticing a mobile layout problem does not establish those properties. When a tool reports that tests passed, we still need to know what was actually tested.

4. Someone still needs to handle problems after automation

Delegating work to tools can change execution without automatically removing costs, failures, or maintenance. Someone needs responsibility for checking schedules, adapting to changed sources, and handling services that stop working.

I no longer compare engineering staff with a partner that “never tires or takes leave.” Programs and models have applicable limits and may need people to investigate errors, add information, and revise workflows. Their value deserves observation, but excluding those burdens before declaring an efficiency multiple is not enough.

A small tool can begin with arrangements that are easier to reverse, change records, and a way to stop use. Whether it is suitable for real work depends on purpose, risk, and actual acceptance, not project size or the number of operators.

5. Four questions to retain before starting

  • Whose problem will be solved, and what counts as completion for this version?
  • Which results have actually been checked, and which are only demonstrations or tool claims?
  • Which tool, labor, service, and maintenance costs are recorded, and which remain uncounted?
  • When failure or changed needs arise, who corrects, takes over, or stops use?

These are the author’s four questions for planning personal AI projects, not proof of engineering competence, team replacement, or a productivity multiple.

6. Start small and leave responsibility with the result

I still want to encourage people to explain a problem, try building, and seek help when stuck. That does not require learning every technology first, nor imply that everyone can independently validate every system.

A small tool with an explainable scope, checkable results, and someone maintaining it deserves serious attention. When expanding later, records can help distinguish what actually helped from what was only an expectation.

Four-panel comic: When AI can help write code, how can someone who has built a small tool know it is usable, rather than merely appearing to run? Whose problem will be solved, and what counts as completion for this version? Which results have actually been checked, and which are only demonstrations or tool claims? Which tool, labor, service, and maintenance costs are recorded, and which remain uncounted? When failure or changed needs arise, who corrects, takes over, or stops use? These are the author’s four questions for planning personal AI projects, not proof of engineering competence, team replacement, or a productivity multiple.

Writing and revision scope: Revised November 4, 2026. This essay develops author reflections from the earlier collaboration questions; it is not a newly verified project case study. Private development logs, quotes, invoices, and historical deployments were not reconstructed. Comparisons involving 12 days, 23,000 lines, 42 person-days, a 3.5-person team, NT$350,000, under NT$3,000, and 117-fold efficiency were removed, along with unverified completed-feature claims about eight platforms, automatic translation, health-fetch frequency, and related charts. Removal does not mean the earlier events necessarily never happened; available evidence is insufficient to republish them as fixed outcomes. The article tool and acceptance scenarios are hypothetical, and the four questions are the author’s suggestions. No new model, software, or productivity experiment was conducted.