Rule reference · Consistency
Contradictory instructions
Fires when: Conflicting formats, lengths, word limits, tools, or languages.
Conflicts usually come from prompts edited over time: someone adds "return JSON" without removing "format the answer in Markdown". The model resolves the conflict differently on different runs.
Keep one instruction, or state which one wins and when ("Return JSON. Use Markdown only inside the summary field").
Example that triggers it
Extract the action items from the meeting notes below. Return JSON with the keys owner and task. Present the final answer in Markdown.
<notes>
{{notes}}
</notes>warning Conflicting instructions: output as JSON and Markdown.
Why it matters. When two instructions conflict, the model picks one unpredictably, so outputs vary between runs.
Fix. Keep one of the two instructions, or say which one wins and when.
Quality 90/100 · 35 tokens · 1 findings
Fixed version
Extract the action items from the meeting notes below. Return JSON with the keys owner and task.
<notes>
{{notes}}
</notes>Quality 100/100 (+10) · 28 tokens · 0 findings
Resolved by the fix
contradictory_instructions: Conflicting instructions: output as JSON and Markdown.
Scores, token counts (GPT-4.1 tokenizer), and findings on this page are computed by the engine when the site is built.
How it affects the score
Each finding subtracts a fixed penalty from 100: critical 30, warning 10, info 3. This rule counts against the Consistency dimension. See the scoring model for the full formula.
Other consistency rules
- Near-duplicate instructions: Two sentences with ≥ 80% word overlap.