PART 01
任务：要做什么 Task: what to do
给出对象和动作。没有任务，候选会分散到闲聊或猜测。 Name the object and action. Without a task, candidates drift toward chat or guesswork.
01 · 同一个同事，不同任务单
01 · Same teammate, different brief
只说“帮我写一下”，和写清任务、禁区、范例，交付会一样吗？ Will “write something for me” produce the same result as a clear task, boundary, and example?
02 提示编辑台：改一段，看候选如何偏转 Prompt Desk: Edit a Section, Watch Candidates Shift
实时输入 live input
任务清晰度
Task clarity
100%
有明确动作
action specified
合规检查
Compliance
PASS
constraint active
格式贴合
Format match
92%
由示例影响
shaped by example
03 三段提示，各管一件事 Three Prompt Sections, Three Jobs
输入结构 input structurePART 02
约束：什么不能做 Constraint: what is forbidden
声明禁区、长度或来源要求；它影响合规候选，但不是权限系统。 State exclusions, length, or source rules. This shapes compliant candidates but is not an authorization system.
PART 03
示例：结果长什么样 Example: what good looks like
少量示例可以示范格式与风格，不会把模型永久训练成这个样子。 A small example demonstrates format and style; it does not permanently retrain the model.
PART 04
候选：只是被影响 Candidates: shaped, not fixed
提示改变当前条件分布，不保证每次生成都绝对服从。 A prompt changes the current conditional distribution; it does not guarantee perfect obedience.
04 最小机制：提示进入条件，不进入权重 Minimal Mechanism: Prompt Enters the Condition, Not the Weights
P(output | prompt)P(y | x, prompt; θ)
θ 是固定模型参数。编辑上方三段只改变
prompt，因此当前 y 的候选分布变化，θ
不变。
θ is the fixed model parameters. Editing the three sections
changes the prompt and current distribution over y, not
θ.
编辑输入
Edit input
任务/约束/示例改变
task/constraint/example changes
条件分布偏转
Distribution shifts
候选权重重新分配
candidate mass reallocates
权重仍固定
Weights stay fixed
下一次仍需带提示
prompt is needed again next time
05 边界实验：删掉敏感信息约束 Boundary Test: Remove the Sensitive-Data Constraint
主动制造失败 Create a failure让“不合规详单”重新抬头 Let the noncompliant detail regain probability
清空约束，观察不合规候选概率上升。提示能影响输出，但真正的敏感信息控制还要靠护栏。 Clear the constraint and watch the noncompliant candidate gain probability. Prompts shape output; real sensitive-data control still needs guardrails.
尚未执行。敏感信息约束当前存在。
Not run yet. The sensitive-data constraint is present.
✗ 写过一次提示，模型就学会了。
One prompt permanently teaches the model.
→ ✓ 提示只在当前推理条件中起作用。
A prompt acts within the current inference condition.
✗ 提示越长，控制越强。
Longer prompts always provide stronger control.
→ ✓ 清楚、无冲突、可验证比堆字更重要。
Clarity, consistency, and testability matter more than
length.
✗ 一句“不得泄露”就是完整护栏。
“Do not leak” is a complete guardrail.
→ ✓ 提示是控制面，不是独立权限或输出拦截器。
A prompt is a control surface, not an authorization or
output-enforcement layer.
06 · 一句话带走
06 · One line to keep