Debug when the model ignores your instructions

Goal: When the model ignores you, do not shout the same boundary louder—shorten the prompt, add one clear example, open a clean thread.

Best for: When the tenth reminder still fails.

Yelling the same Boundary in ALL CAPS rarely helps. Reset the chat, put five lines on top, show one example.

Models do not “obey” like employees. Long chats bury early rules. Conflicting instructions cancel each other. Vague vibes never bind. Repair the setup.

When the model ignores a rule, do not shout louder—isolate the rule, put it first, add an example, and cut conflicting instructions.

Change one thing per retry: order, example, or cut conflicts. Changing five things teaches you nothing.

Why beginners skip this

Swapping models when it “ignores you” feels tidy—the real break is conflicting constraints. Split the ask, then change.

Step 1 — Common causes

• Rules buried under a huge paste
• Conflicting rules in one message (“be brief” + “cover everything”)
• Chat too long—early rules fell out of usable context
• Rule was vague (“be professional”)
• You rewarded bad output by continuing instead of resetting

Step 2 — Repair sequence

Fix in this order: use a real failed sample—do not invent a pile of “MUST obey.”

1) New chat (or fresh chat inside the same Project)
2) Put rules at the top in ≤5 lines
3) Add one short EXAMPLE of correct output
4) Ask the model to restate rules before answering
5) Only then paste the real task

My prompt says “no emojis” but replies still have them. Write a debug sequence: extract rules, reorder, add bad/good example, shrink prompt. Show the revised prompt for a polite decline email.

Step 3 — If it still drifts

Move standing rules into Project instructions. Add a rubric checkbox. Shorten the user message—huge pastes drown Constraints.

Live task this week (not a toy): apply this guide's Goal — Fix ignored rules with shorter prompts, a clear example, and a clean new chat—instead of yelling the
Fill with your real artifact names/files. Constraints: no invented facts; label Unknowns; stop before send/delete/pay.
Return the working prompt + one pass/fail check I can mark in a minute.

Save this card

Card: When the model ignores you: extract the rule, reorder, give a counterexample, then shorten. Keep this beside the tool; read before you start.

Before you start

Keep the last failing output visible. You will repair the prompt for “Fix ignored rules with shorter prompts, a clear” against that failure—not a fresh toy ask. Guide: prompt-debug-when-it-ignores-you. Keep the real job visible the whole time; a blank box means stop and fetch the artifact first.

How to use this week

This week for one real ignore: split goal/format/bans into three lines, delete conflicts, then rerun.

One change per retry

Reorder, add example, or cut conflicts—not all three at once.

Keep it real

Run this guide’s Practice now on work you will actually ship this week—not a fictional scenario. Dated notes beat perfect theory.

Ship the practice

Close the tab only after Practice now is done on a real artifact.

Do not bookmark-only

Finish Practice now, then tick this guide on your path. Titles unread are not skills.

Close the loop

Write one line: next time I swap models without splitting the ask, I refuse.

Check the unknown path

What should the model do when a field is missing—null, TBD, or ask you? Write it once in Boundaries.

Model ignored my "max 80 words" twice.
Debug card: restate constraint, ask it to quote the constraint, then regenerate.
Log which fix worked.

Red flag

Repeating the same ignored rule louder; never extracting/reordering—use the four-step debug.

Pitfalls

Temptation: skip the habit for prompt-debug-when-it-ignores-you to save five minutes.
Job still: Fix ignored rules with shorter prompts, a
Smallest correct pass: Done visible · checkable artifact · Unknowns [?].
Boundaries: draft only; no silent send; no invented metrics.
Return: artifact + refuse-line said aloud.

Practice now

Priority stack: rank constraints 1..n; put #1 first (Asia/Shanghai).

Constraint census: number the constraints; ask model to echo the list (Asia/Shanghai).

Bold-boundary drill: move the ignored rule to line 1 in ALL CAPS refuse form (Asia/Shanghai).

Ignore-debug drill: quote the ignored constraint; rewrite it as a hard boundary line (Asia/Shanghai).

Thin-spot drill: rewrite the Save card with real nouns from your work (repo, doc, inbox). Generic placeholders fail the drill.

Tonight for one real ignore split goal/format/bans into three lines, delete conflicts, then rerun.

One sentence takeaway

Split goal/format/bans three lines; delete conflicts, then rerun.

Next

GFB · Few-shot · Context windows

Floor 0336b: one dated Asia/Shanghai refuse-line on the Save card before quit.

模型无视你的说明时如何调试

目标： 模型老是无视你的说明时，别把同一条边界吼得更大声。换更短的提示、一个清楚例子、一次干净的新对话。

适合： 同一条规则提醒到第十次，还是失败的人。

全大写很少管用。更稳的做法：新开对话，把规则压到最上头五短行，再给一个例子。

模型不会像员工那样「听话」。长对话会把早先规则埋掉；互相打架的说明会抵消；「专业一点」这种感觉词没有约束力。该修的是设置，不是嗓门。

每次重试只改一样：调顺序、加例子，或删掉冲突句。一次改五样，你什么也学不到。

为什么新手跳过

模型「不听话」就换模型像省事；真正坏在约束互相打架。先拆指令再改。

步骤 1——常见原因

• 规则埋在大段粘贴下面
• 同一条消息里规则冲突（「简短」又「面面俱到」）
• 对话太长：早先规则离开有效上下文
• 规则太含糊（「专业一点」）
• 你用继续聊奖励了糟糕输出，而不是重置

步骤 2——修复顺序

按下面顺序修：材料用真实失败样例，别凭空堆「务必遵守」。

1）新对话（或同一 Project 内开新聊天）
2）把规则放在最顶，≤5 行
3）加一个短的正确输出 EXAMPLE
4）先让模型用自己的话复述规则
5）然后再贴真正任务

提示词写了「无表情」却仍出现表情。请写调试序列：抽取规则、重排、加坏例/好例、缩短提示；并展示用于礼貌拒绝邮件的修订提示词。
边界：不编造；未知标 [?]；发送前停下。
完成检查：一分钟内能判过/不过。

步骤 3——如果仍然漂移

把常驻规则移进 Project 说明，加上评分勾选，再缩短用户消息。大段粘贴很容易把约束淹死。

列出模型「无视」长度上限的 5 个常见原因，以及各自怎么改。
背景：这是本周真实工作。要求：不编造；只出草稿；写给新手也能看懂。
边界：不编造；未知标 [?]；发送前停下。
完成检查：一分钟内能判过/不过。

保存这张卡

记下：今天改管用的那条提示词（含边界）+ 版本或日期。

开始前

把上一次失败输出留在眼前。没有失败样例的调试只是猜。

本周怎么用

本周对一次真忽略：把目标/格式/禁令拆成三行，删打架句后再跑。

每次重试只改一样

重排、加例子、或删掉冲突指令：一次只动一样，不要三样一起改。

保持真实

用本周真要交付的工作做「现在就练一次」，别编假场景。带日期的实操笔记，比完美理论有用。

把练习做完

用真实材料做完「现在就练一次」，再关标签页。

不要只收藏

先做完「现在就练一次」，再在学习路径上打勾。只收藏不练，标题变不成技能。

闭环

写下一行：下次不拆指令就换模型，我拒绝。

检查未知路径

字段缺失时模型该怎么做：填 null、写 TBD，还是问你？在边界说明里写清楚一次。

红旗

模型「不听话」就换模型，真正坏在约束互相打架；先拆指令再改。

常见坑

现在就练一次

今晚就练：对一次真忽略把目标/格式/禁令拆成三行，删打架句后再跑。

一句话带走

目标/格式/禁令拆三行；删打架句后再跑。

下一步

目标·格式·边界 · 少样本 · 上下文窗口