讓 AI 工作流程有空間犯有用的錯誤

“Make no mistakes” sounds like a serious instruction. As a development process, it leaves out the most useful part: what happens when an assumption turns out to be wrong. A working prototype can expose that assumption in minutes. A longer prompt demanding perfection can keep it hidden behind a more confident answer.

Disclosure: An AI agent drafted this article from my direction, personal background, public project documentation, and the linked technical sources.

I started building AI-powered tools in 2019, before ChatGPT's public launch. The habit I want to carry into today's tools is simple: try something concrete, inspect what happened, and let the next attempt benefit from the mistake. Humans improve that way. An AI workflow needs a deliberate mechanism for carrying the lesson forward.

Give the experiment a question

There is a difference between asking for “a complete, flawless application” and asking whether a person can find one useful record and understand it. The second request creates something you can evaluate. It also leaves room to discover that the search box, the record structure, or your original idea needs to change.

My project archive spans application tools, voice workflows, supplier pricing, archive retrieval, community platforms, and computer-vision experiments. These are different kinds of work, with different maturity levels. I do not expect one enormous instruction to settle all their product and engineering questions in advance.

For a new idea, I would start with one uncertainty: Can the input be interpreted? Can the proposed interaction be understood? Can the output survive the next step? Then build the smallest attempt that produces evidence about it. A result that disproves the idea can be valuable if it prevents a week of polishing the wrong thing.

Make the mistake cheap to recover from

Room to experiment works best when the attempt has a boundary. Use a branch, a recoverable draft, sample records, or a practice environment. Set a concrete check before the experiment begins. A failed layout in a preview is a different situation from silently publishing that layout to customers.

Okeike's public case study describes recoverable drafts, inspectable AI-proposed edits, and publishing through versioned source control. Exempliph.ai's case study describes a practice workspace and a distinction between prepared applications and confirmed submissions. Those are documented design choices, not evidence that either product makes mistakes impossible.

For a comparable editing experiment, I would ask for a few navigation alternatives in a draft, try them against an actual page, and keep the one that makes the next action clearer. The discovery work can be flexible while publication remains a separate decision.

That balance gives creativity somewhere to operate. The model can suggest an unexpected structure without also receiving permission to overwrite every page. You can explore widely without making every experiment expensive to undo.

Turn the failed attempt into a reusable lesson

“It did not work” is too vague to improve the next run. Keep the input that exposed the issue, the expected behavior, the observed result, and the change that fixed it. When appropriate, turn that example into a regression check.

For example, a form mapper might confuse an employer's name with a job title. The useful correction is a specific mapping example and a check that preserves an existing answer. Adding “never misunderstand any form” to the prompt does not capture the distinction the failure revealed.

This is also where the human analogy has a limit. An ordinary inference session does not automatically retrain a model when you correct it. The workflow learns operationally when you retain the evidence in examples, tests, instructions, or retrievable notes. If the next run never receives that lesson, you may pay to rediscover it.

Anthropic's September 29, 2025 context-engineering article recommends starting with a minimal useful prompt, adding guidance from observed failures, and keeping context informative. It also describes persistent notes for long tasks. This is established engineering guidance, not a new October release.

Measure the whole attempt, not the length of the prompt

Allowing iteration can reduce token use when it replaces repeated broad rewrites, bloated instructions, or the same investigation performed again. It can also increase token use if the agent wanders, repeats failures, or generates ten alternatives when one would answer the question. Exploration is not a guaranteed discount.

Compare completed outcomes. Record model usage across the initial attempt, inspection, repairs, and final verification. Keep an eye on the human review time too. A short prompt with endless retries can be worse than a longer prompt containing the right example.

My preferred next instruction is specific: keep the working parts, change the failed behavior, and explain which evidence shows the correction worked. Once an approach is understood, ordinary code or a focused check may handle it more economically than asking a model to reconsider it every time.

I bring this attitude into the seminars and practical sessions I host. At AFROTECH 2026, my Building AfroTech.guide session will walk through a particular project and offer practical steps people can use in their own builds. The point is to make the process understandable, including the decisions that need another attempt.

A useful mistake changes the next experiment. Give the system a clear question, a recoverable place to try, and a way to retain the result. That creates more room for original ideas while keeping progress grounded in something you can inspect.