Understand what the engine checks

min-length — specificity

Prompts shorter than 20 words rarely give a model enough to work with. The score scales linearly with the word count below the threshold.

Fix: Add detail about what you want, why, and how the output should look.

Rewrite (append): Flesh out the prompt

<context>
  <who is asking and why>
</context>

<instructions>
  <what you want, step by step>
</instructions>

<output_format>
  <exact shape of the answer>
</output_format>

max-length — structure

Very long prompts (>1500 words) tend to contain redundancy and dilute the model’s focus. The score decreases gradually past the soft limit.

Fix: Look for repeated instructions, bundled unrelated tasks, or sections that can be summarized.

no-output-format — specificity

The prompt should tell the model exactly how to format its answer, otherwise the model will guess and consumers cannot rely on the shape.

Fix: State the exact format: JSON schema, bullet list, markdown table, single sentence, etc.

Rewrite (append): Specify the output format

<output_format>
  Return a <JSON object | bullet list | markdown table | single paragraph>.
  Required fields / sections: <list>.
  Do not include: <what to omit>.
</output_format>

no-examples — best-practice

Examples dramatically improve consistency on classification, extraction, and formatting tasks. Profiles like claude weight this rule higher.

Fix: Add 1–3 concrete examples showing the input and the expected output.

Rewrite (append): Add few-shot examples

<examples>
  <example>
    <input><sample input></input>
    <output><expected output></output>
  </example>
  <example>
    <input><sample input></input>
    <output><expected output></output>
  </example>
</examples>

no-role — best-practice

Assigning a role focuses the model and sets expectations for expertise and tone. Useful for both system and user messages.

Fix: Start with something like "You are a senior <X> who specializes in <Y>."

Rewrite (prepend): Assign a role

You are a senior <role> who specializes in <domain>. You write for <audience> and prioritize <quality bar>.

no-context — specificity

Context helps the model understand the situation, audience, and constraints. Long prompts (>=80 words) are assumed to provide implicit context.

Fix: Explain the situation: who the user is, why they’re asking, and what the stakes are.

Rewrite (prepend): Add a context block

<context>
  Who is asking: <user>
  Why they are asking: <motivation>
  Constraints or stakes: <what cannot break>
</context>

ambiguous-negation — clarity

Models follow positive instructions ("do Y") more reliably than negations ("don't do X"). Heavy stacking of negations correlates with regressions.

Fix: Rewrite "don't do X" as "do Y instead". Tell the model what the desired behavior is.

no-constraints — specificity

Constraints keep the model on track and prevent scope drift. They are also the easiest hook for downstream evaluation.

Fix: Add constraints like length limits, scope boundaries, or things the answer must include.

Rewrite (append): Add explicit constraints

<constraints>
  - Length: <e.g. ≤ 200 words>
  - Scope: <what is in scope and what is not>
  - Must include: <required elements>
  - Must avoid: <forbidden elements>
</constraints>

all-caps-abuse — clarity

Excessive ALL CAPS is noisy and rarely the clearest way to emphasize something. Bold, quotes, and XML tags work better.

Fix: Use bold (**word**), quotes, or XML tags for emphasis instead of ALL CAPS.

vague-instruction — clarity

Vague qualifiers ("good", "proper", "appropriate") don't give the model a measurable target. Replace them with concrete acceptance criteria.

Fix: Replace vague words with measurable criteria. "Good" → "concise (≤ 3 sentences) and citing sources".

missing-task — clarity

This is the highest-weight rule and the only one that fires as an error by default. If the model can't identify a task, the rest of the prompt is wasted.

Fix: State the task explicitly: "Your task is to..." or "Please <verb> <object>".

Rewrite (prepend): Add an explicit task

Your task is to <verb> <object>. Specifically: <what success looks like>.

no-structured-format — structure

Long prompts (>100 words) are easier for a model to follow when broken into sections. XML tags work especially well for Claude; markdown headers work well for GPT.

Fix: Split the prompt into labeled sections: <instructions>, <context>, <examples>, <output_format>.

Ambiguity

The prompt has multiple plausible readings, missing scope, or under-specified inputs.

Fix: Specify the task, input, audience, constraints, and expected output.

Rewrite (append): Make the ambiguous parts explicit

<inputs>
  <input variable> = <where the data comes from>
</inputs>

<scope>
  In scope: <what to consider>
  Out of scope: <what to ignore>
</scope>

<expected_output>
  Shape: <exact format>
  Audience: <who is reading>
</expected_output>

Conflicting instructions

The prompt asks for incompatible behaviors (e.g. JSON and plain text) or contradicts itself across sections.

Fix: Choose one instruction path and remove the incompatible wording.

Rewrite (append): Pick one consistent set of instructions

<resolved_instructions>
  Output format: <ONE choice — JSON | markdown | plain text>
  Length: <ONE target — exact words / characters / sentences>
  Tone: <ONE register — formal | casual | technical>
</resolved_instructions>

Remove any earlier wording that contradicts these resolved instructions.

Grounding

The prompt expects facts, citations, or jurisdiction-specific knowledge without supplying source material or scope.

Fix: Provide source material, scope, assumptions, and how the model should handle uncertainty.

Rewrite (append): Add the missing grounding

<grounding>
  Sources: <attached docs, URLs, or {{variable}} placeholders>
  Scope: <jurisdiction, time period, domain>
  Assumptions: <what the model can take as given>
  Uncertainty: If the source material is silent on a point, say "not stated in the provided material" rather than guessing.
</grounding>

Success criteria

The prompt does not state what a good answer must include, avoid, or optimize for.

Fix: State what a good answer must include, avoid, and optimize for.

Rewrite (append): Define measurable success criteria

<success_criteria>
  A good answer must include: <required elements>
  A good answer must avoid: <forbidden elements or patterns>
  Optimize for: <one explicit metric — accuracy | brevity | citation density | etc.>
</success_criteria>

Task framing

The task is unrealistic, overbroad, or asks for guarantees the model cannot reasonably provide.

Fix: Narrow the task to something the model can reasonably complete and verify.

Rewrite (prepend): Reframe the task into something verifiable

Narrow the task to a single deliverable: <one concrete output>.
Out of scope: <items the model should not attempt>.
A reviewer can verify success by checking: <one observable check>.