The Prompt Lab — Salience Weighting
Learn the salience weighting prompting technique with concrete before/after examples.
Salience Weighting
The Technique
Salience weighting means explicitly telling the model which parts of your request matter most — and by how much — so it allocates its “attention” accordingly. Without this, models treat all requirements as roughly equal, often optimizing for easy wins (length, formatting) while underserving the harder constraints that actually matter to you.
The Naive Prompt
Write a subject line for a cold email to a VP of Engineering at a mid-size SaaS company.
We're selling a developer observability tool. Keep it short, make it personal,
avoid sounding spammy, and make it curiosity-driven.Why It Falls Short
Four constraints, zero hierarchy. The model has no way to know that “avoid sounding spammy” is a dealbreaker while “make it personal” is a nice-to-have, so it splits the difference and produces something generically inoffensive. You often get competent mediocrity — a subject line that technically satisfies all four criteria weakly rather than nailing the one that matters most.
The Improved Prompt
Write a subject line for a cold email to a VP of Engineering at a mid-size SaaS company.
We're selling a developer observability tool.
Prioritize these constraints in order — trade off lower ones to serve higher ones if needed:
1. [CRITICAL] Curiosity-driven — the recipient must feel genuinely compelled to open it
2. [HIGH] Under 8 words — mobile preview truncates anything longer
3. [MEDIUM] Sounds like it came from a peer engineer, not a marketer
4. [LOW] Avoids generic phrases like "quick question" or "following up"
Generate 5 options. After each one, note which constraint you leaned into most heavily.Why It Works
Labeling constraints as CRITICAL, HIGH, MEDIUM, and LOW gives the model a clear tiebreaker when requirements conflict — curiosity wins over peer tone if it has to choose. Asking the model to annotate its own reasoning after each option also surfaces trade-offs you’d otherwise never see, letting you select or remix intelligently rather than just picking blindly from a list.
When to Use This
- When your prompt has 4+ requirements and some genuinely matter more than others — common in marketing copy, legal summaries, and product spec writing
- When you’re getting outputs that feel “safe but flat” — salience weighting often unlocks bolder responses because the model knows which risks are worth taking
- When working with high-stakes, cost-sensitive tasks on a capable model like Claude Fable 5.1 or GPT-6 Astra, where you’re paying for nuanced judgment and want the model to actually exercise it rather than hedging across all dimensions equally