Most prompt failures come from a small set of repeated mistakes. Each one below has a symptom you will recognize, a fix, and a link to the PromptFlowEngine rule that detects it, with a triggering example and a corrected prompt.
Clarity mistakes
1. No stated objective
The prompt describes material and never says what to do with it. The model guesses between summarizing, analyzing, and rewriting.
Fix: open with an imperative verb that names the outcome. Reference: missing objective.
2. Vague language
Words such as “something”, “stuff”, “things”, and “etc.” ask the model to invent the specifics.
Fix: list the actual items. If the list is open-ended, bound it: “up to 3 of these”. Reference: vague language.
3. Hedged instructions
“Try to”, “if possible”, and “maybe” turn a requirement into a suggestion, and the model treats it as optional.
Fix: state it as a requirement, or say exactly when it applies. Reference: hedged instruction.
4. Only negative constraints
A prompt made of “don’t” statements tells the model what to avoid and nothing about what to produce.
Fix: pair each prohibition with the behavior you want. Reference: negative-only constraints.
Specificity mistakes
5. Sizes without numbers
“Short”, “a few”, and “detailed” vary between runs, which breaks layouts and parsers.
Fix: use a number and a unit, such as “under 150 words”. Reference: unquantified size.
6. No output format
Without a format, the model chooses between prose, bullets, and a table, and chooses differently each time.
Fix: specify the structure, and the schema if code reads the result. References: missing output format and missing length guidance.
7. References to things that are not there
“Fix the bug above” or “use the attached file” in a prompt that contains neither.
Fix: include the material, or name where it will be inserted. Reference: unclear reference.
Consistency mistakes
8. Contradictory instructions
“Keep it short but be detailed.” Long prompts edited by several people collect these. The model resolves the conflict on its own, unpredictably.
Fix: decide the priority and say it: “Be concise. Where brevity and accuracy conflict, choose accuracy.” Reference: contradictory instructions.
9. Repetition
Saying something twice for emphasis costs tokens. Saying it twice with slightly different wording creates ambiguity about which version counts.
Fix: state each instruction once, in the section where it belongs. References: duplicate instruction and near-duplicate instruction.
Structure mistakes
10. A wall of text
One long paragraph that mixes task, context, rules, and input. Instructions buried in the middle get less weight, and the next editor cannot find anything.
Fix: split it into Task, Context, Requirements, Output format, and Input. Reference: wall of text.
Security mistakes
11. Undelimited user input
A template variable placed directly in the instructions lets whatever the user types read as part of them.
Translate this to French: {{user_text}}
Fix: wrap the variable in tags and tell the model the tags contain data. Reference: undelimited variable. Background: prompt injection.
12. Secrets and personal data in the prompt
An API key in a pasted config, a password in a log, a customer email in a ticket. Once sent, it is in a third party’s systems.
Fix: redact before sending and keep credentials out of prompt files. Reference: sensitive data. Background: PII and secret detection in prompts.
Process mistakes that rules cannot catch
- Testing on one input. A prompt that works on your example may fail on real traffic. Build a small test set. See prompt evaluation.
- Changing several things at once. When the result changes, you will not know why. Change one thing per iteration.
- Not versioning prompts. If a prompt lives only in a chat window, you cannot roll back. See prompt management.
- Stuffing the context window. More context is not always better. See token optimization.
Find them in your own prompts
Paste a prompt into the PromptFlowEngine prompt analyzer. It runs every rule above, quotes the text that triggered each finding, and suggests a fix. The full rule catalogue lists all of them.