The Prompt Lab — Priority Filtering
Learn the priority filtering prompting technique with concrete before/after examples.
Priority Filtering
The Technique
Priority Filtering is the practice of explicitly telling the model which dimensions of quality matter most — and in what order — before it begins generating. Models like GPT-5.6 Sol and Claude Opus 5 optimize across many implicit objectives simultaneously (coherence, completeness, tone, brevity); without a stated hierarchy, they guess. Declaring priorities collapses that ambiguity and steers generation toward your tradeoffs rather than the model’s defaults.
The Naive Prompt
Write a follow-up email to a client who missed our onboarding call
and hasn't responded to two previous emails.Why It Falls Short
The model has no idea whether you want warm or firm, brief or thorough, relationship-preserving or deadline-driven. You’ll typically get a blandly diplomatic middle-ground email — inoffensive, forgettable, and unlikely to get a response. Every constraint is weighted equally, so nothing is weighted meaningfully.
The Improved Prompt
Write a follow-up email to a client who missed our onboarding call
and hasn't responded to two previous emails.
Prioritize in this order:
1. BREVITY — under 80 words, no exceptions
2. SPECIFICITY — name the concrete consequence (their account setup
stalls after Friday if we don't connect)
3. WARMTH — assume good intent; they're probably just swamped
Do NOT: apologize for following up, use "just checking in,"
or offer to reschedule without a specific time already proposed.
Proposed time: Thursday, August 6 at 2pm ET or Friday, August 7 at 10am ET.Why It Works
The numbered priority list tells the model how to resolve conflicts — when brevity and warmth fight for the same sentence, brevity wins. Naming the consequence gives the model a specific anchor instead of a vague urgency signal, and the explicit exclusions cut the three most common failure patterns before they appear. The output becomes a decision, not a draft.
When to Use This
- When multiple qualities are genuinely in tension — you want an analysis that’s both comprehensive and skimmable, or feedback that’s honest and encouraging. Without a tiebreaker, models hedge.
- For high-stakes single-shot outputs — pitch decks, client-facing copy, legal summaries — where you don’t want to iterate through five drafts before the model understands your actual tradeoffs.
- When you keep editing the same dimension repeatedly — if you always shorten the output, always punch up the urgency, or always remove the qualifications, that’s a sign your implicit priority isn’t being communicated; make it explicit.