Master AI Coding with a 7‑Step Loop: From Task Cards to Release Checks

Introduction

Why does AI‑generated code feel great for the first half hour and then require a lot of bug‑fixing? The problem is not the model but the use of a one‑off prompt that tries to pack requirements, design, implementation, and testing into a single sentence.

Prompt is not dead, but one‑off prompts should retire

Instead of treating a prompt as a wish‑list, it should become the entry point of a repeatable loop. The article contrasts a one‑off prompt:

帮我给订单列表加一个导出 Excel 功能，要求支持勾选订单、没有勾选时给提示、数据多的时候不要卡。

with a Loop Engineering prompt that lists the steps “需求说明 → 澄清问题 → 拆任务 → 写测试 → 写代码 → 跑验证 → 保存进度”.

What a Loop looks like

The minimal Loop consists of four artefacts: 需求.md – a requirements document that defines what to do, what not to do, and the acceptance criteria. 任务.md – a task list that breaks the large requirement into 3‑5 small steps. 验证.md – a verification record that logs which tests were run and their results. 进度.md – a progress snapshot that records what has been completed, what remains, and the next entry point.

The four artefacts must satisfy four invariants: bounded requirements, ordered tasks, evidence of completion, and resumable state.

Core principle: turn “I want” into “You should do”

One‑off prompts often say “I hope the code quality is high”. Loop prompts replace such vague wishes with concrete clauses and acceptance criteria, e.g.:

需求条款：空选提示

当用户没有勾选任何订单时：
- 系统不能开始下载文件
- 系统要提示「请先选择要导出的订单」

This makes the AI’s work verifiable, testable, reviewable, and resumable.

Daily workflow: 7 steps

Step 1 – Create a task card

Give the task a simple name (e.g. “订单导出功能”) and ask the AI to produce a task card containing goal, constraints, “not doing” items, and acceptance standards.

目标：用户可以导出自己勾选的订单
约束：
- 没勾选订单时不能导出
- 数据很多时页面不能长时间无响应
- 导出的文件要能被 Excel 正常打开
不做：
- 暂时不做后台异步导出
- 暂时不做自定义字段选择
验收标准：
- 勾选 1 条订单，可以导出 1 条
- 没勾选订单，会出现提示
- 特殊字符不会把表格格式弄乱
- 跑完测试后才能算完成

Step 2 – Clarify ambiguous words

Identify vague terms such as “大数据量”, “不要卡”, “兼容 Excel”, “导出当前数据” and turn them into concrete, checkable conditions. Example mapping:

大数据量 → >10000 条时显示进度

不要卡 → 分块 serialize，并有状态更新

兼容 Excel → UTF‑8 with BOM

导出当前数据 → 仅导出当前勾选的订单

Step 3 – Write a plan instead of jumping straight to code

Break the task into 3‑5 small sub‑tasks, each with a clear impact area and acceptance criteria. Example sub‑tasks for the export feature include file generation, button interaction, data selection, and large‑data handling.

Step 4 – Run a TDD mini‑loop

For each sub‑task execute “fail test → minimal implementation → test passes → update task status”. Example test for export:

1. 先写一个测试：订单名称里有逗号、换行、双引号时，导出内容仍然正确
2. 跑测试，确认失败来自功能未实现
3. 写导出函数
4. 再跑测试
5. 通过后进入下一个任务

Step 5 – Record evidence, not just “done”

A verification record must contain the exact test command and key results, e.g.:

测试命令：npm test
结果：全部通过，23 个测试无失败

覆盖率命令：npm run test:coverage
结果：分支覆盖率 86%，达到 80% 的最低要求

Step 6 – Save a progress snapshot

Before interrupting, write a snapshot that lists completed items, pending items, evidence, and the next entry point. Example snapshot shows completed task card, task breakdown, export function, empty‑selection prompt, and passed tests, and lists next steps such as final acceptance check and security scan.

Step 7 – Perform a release check

Compare the release checklist against the task card: all acceptance criteria satisfied, all automated tests passed, no new security risks, no secrets in code, and a progress snapshot exists. The result is either PASS (merge or deliver) or BLOCKED (list blockers).

Five sentences to remember

Start a new task with a task card, not direct code.

Clarify ambiguous terms in at most three rounds, marking confirmed and pending items.

Execute the TDD mini‑loop: fail test → minimal implementation → test passes → update progress.

When pausing, write a progress snapshot that includes evidence and the next entry phrase.

Before closing, run the release checklist and only proceed if it PASSes.

When to use a full Loop

Three levels of rigor are suggested:

Lite – direct fix with verification (e.g., typo correction).

Standard – task card, task list, tests, and progress snapshot (e.g., new page interaction).

Full – adds branch isolation, code review, and security gates for high‑impact changes (e.g., new module, payment integration).

If a change could affect users, data, security, or collaboration, avoid a one‑off prompt and adopt at least the Standard Loop.

Common pitfalls

Pitfall 1: Only naming the process

Without a task card, progress snapshot, or verification record, you are still using a prompt.

Pitfall 2: Tests without acceptance criteria

Passing tests does not guarantee the requirement is met; always cross‑check against the original acceptance standards.

Pitfall 3: Treating AI promises as hard rules

Only enforceable artefacts are test commands, coverage thresholds, security scans, release checks, and human review.

Pitfall 4: Forgetting a resumable entry

When a task is paused, add a line such as “请读取进度快照，从‘下一步’继续”。

A minimal copy‑paste template

按 Loop Engineering 工作流处理这个任务：

目标：
<一句话说明要做什么>

约束：
- 不改无关文件
- 不提交密钥
- 先写任务卡，确认后再实现
- 每一步都要更新任务清单或进度快照

流程：
1. 写任务卡：目标、约束、不做什么、验收标准
2. 找出模糊词，先问清楚
3. 拆成 3-5 个小任务
4. 每个小任务写清楚验收口径
5. 按 TDD 小循环实现：失败测试 → 最小实现 → 测试通过
6. 跑测试、覆盖率、安全检查
7. 对照验收标准逐条检查
8. 写进度快照，留下续跑入口
9. 做放行检查，PASS 才能合并或交付

退出条件：
- tests PASS
- coverage 达标
- security PASS
- 放行检查 PASS
- 进度快照可用于下次续跑

Summary

Prompt engineering solves “how to say it this time”; Loop Engineering solves “how to do it every time”. By making the requirements, task list, tests, verification, progress snapshot, and release checklist explicit, AI‑generated code becomes observable, recoverable, and reliably deliverable.