Gains:
- Recognizes the most common prompt errors from real examples
- Can apply a practical and repeatable fix for every error
- Realizes that bad output is often caused by a bad prompt
In this module, we learned the techniques of writing good prompts one by one. Now let's reinforce these from a different angle: by looking at the most common mistakes and their solutions. The vast majority of bad AI output is due to bad prompting, not model inadequacy. If you recognize errors in this unit, you can quickly diagnose and fix problems in your own prompts; Instead of saying "AI can't do this" you can say "let me fix my prompt like this".
Why Is It Important to Recognize Errors?
When an output does not meet your expectations, you can react in two ways: blame the tool or question the prompt. The experienced user does the latter. Because most of the time, the source of the problem is obvious and the result changes with a small correction. Below we will see the eight most common errors, their symptoms and solutions.
The Eight Most Common Mistakes
Error
symptom
Solution
ambiguity
Generic, "one size fits all" answer
Give measurable instruction
Zero context
Output that does not fit the situation
Add who/why/history
Not specifying format
Output in wrong format
Force format
absence of constraints
Excessive, deviant content
Add a "don't" list
overload
Model skips some requests
Divide work, prioritize
non-confirmation
Using fake information
Check facts with source
One shot expectation
Don't give up at the first exit
Iterate
contradictory instruction
inconsistent output
Align constraints
Step by Step Diagnostic Method
When a corrupted output arrives, follow this order:
- Is the mission clear? Is the verb concrete, does it mean the only thing you want?
- Is context enough? Does the model know the situation?
- Has the format been mentioned? Is the format the way you want it?
- Are there any restrictions? Is it written what is not wanted?
- Did I load too much? Did I ask for 5 different jobs in one prompt?
- Did I verify? Have the number, name, date been checked?
These six questions catch almost all errors.
Four Copyable Proofreading Templates
1) Removing ambiguity:
Make this instruction measurable: [vague instruction]Translate each vague point into a concrete number, quantity, or criterion.
2) Splitting the overload:
Don't do this task all at once. Proceed in this order and stop when you finish each step, wait for my confirmation:1) [subtask 1]2) [subtask 2]3) [subtask 3]
3) Limiting fitting:
Just rely on the information I provided below. If there is no information, write "no data"; do not guess or add. Indicate which line each important claim you use comes from. Source: [text]
4) Resolving the contradiction:
Check if there are any contradictions in the instructions I gave you. If so, list the conflicting points and ask me which one to prioritize; then produce accordingly.
Weak Prompt / Strong Prompt
Weak (many errors combined):
Write a comprehensive plan for our company about marketing, budget, and recruiting, and you're good to go.
Powerful (bugs fixed):
Role: You are a business development specialist advising SMEs. Context: A software company with 12 people; goal is 20 new customers this quarter.Task: Write marketing plan ONLY (do not include budget and hiring).Format: 5-point action list; each item: action + responsible + criteria.Constraint: Paid advertising budget 0; Recommend only organic channels. Don't give fake metrics; Justify the suggestion.
The powerful version splits the overload (cramming three topics into one prompt), replaces ambiguity ("good luck") with criteria, and adds formatting and constraint.
Three Mini Cases
Case 1 — Overload. A manager requested a report summary, presentation plan and e-mail in a single prompt; The model made all three in half. By dividing the job into three separate prompts, each output was complete and usable. Total time was less than struggling with a single complex prompt.
Case 2 — Non-confirmation. A content team published an “industry statistic” given by the model without verifying it; The number turned out to be wrong and they had to issue a correction. Afterwards, they adopted the rule of "give each number with its source, I will check the source" and eliminated this risk.
Case 3 — Conflicting instruction. When a marketer said "it's very short but tell all the features", the model skipped some features. When I made the instruction non-contradictory as "describe the 3 most important features in 60 words", the output was both short and complete.
Tip: Run your prompt through the “six diagnostic questions” before going live. This 30-second check makes most rounds of iteration unnecessary from the start.
Caution: The most dangerous mistake is "non-validation" because it is overlooked because the output looks smooth and convincing. The model can match a number, name or source with great confidence. Independently verify any fact to be published or used as a basis for decision. We will deepen this topic in the next unit.
Invisible Error: "Almost Correct" Output
Some errors are easy to catch: the output comes in the wrong format, goes off topic, or comes up blank. But the most insidious mistake is that the output is almost correct. The text is fluent, the structure is smooth, the tone is on point; There's just one wrong number, an overlooked constraint, or a small logical error. These types of outputs are dangerous because they escape inspection because they "look good."
Two habits protect against this. First, do not use the output without reading it: no matter how rushed you are, read it through before sending it. Second, having the model check its output: “check if you followed all the constraints I gave you” or “list all the numbers in this text” will bring hidden errors to the surface. These two steps take a few seconds, but the cost of incorrect decisions caused by “almost correct” output is much higher.
Avoiding mistakes: design over control
Instead of hunting for errors in the output, the experienced user tries to prevent them from the beginning at the prompt. This is a difference in mentality. Turn each recurring error into a "permanent rule": if the model keeps adding comments in a summary job, now set the "don't add comments" constraint to every summary prompt. If the model always produces too many items in a list job, place the "exactly X items" constraint in your template. In this way, instead of catching the same error again and again, you prevent it from occurring in the first place. A well-designed prompt replaces dozens of subsequent corrections; This turns error diagnosis from a burden into a learning and improvement tool.
Common mistakes
- Blaming the middleman. Saying "AI can't do it" instead of looking for the problem in the prompt.
- Putting too much work into a single prompt. Compressing five separate tasks into one instruction and doing them half-done.
- Skipping verification. Assuming the fluent output is correct.
- Giving contradictory constraint. Self-conflicting demands, such as “short but comprehensive.”
- Trying again without diagnosing. Repeating the same mistake without finding out what's broken.
In summary
- Most bad output is a problem with the prompt, not the model; can be diagnosed and corrected.
- The most common errors: ambiguity, zero context, lack of format, lack of constraints, overload, lack of verification, one-shot expectation, and contradiction.
- Six diagnostic questions (task, context, format, constraint, load, verification) catch most errors.
- Dividing complex jobs gives faster and higher quality results than loading them onto a single prompt.
- The most insidious error is non-verification; Smooth output does not mean accurate.
Application task
Find an old printout you don't like and run its prompt through the "six diagnostic questions." Mark which errors occur, apply the fix in this unit for each, and run the prompt again. Note which single correction produced the greatest improvement, not how many rounds.
checklist
- [ ] In case of bad output, I query the prompt first, not the agent.
- [ ] I can perform six diagnostic questions.
- [ ] I divide complex tasks into multiple prompts.
- [ ] I notice conflicting and ambiguous instructions and correct them.
- [ ] I verify the factual output before publishing it.