The Prompt Lab — Failure Mode Injection

Failure Mode Injection

The Technique

Failure Mode Injection means explicitly telling the model what not to do — and why those failure modes are tempting — before asking it to produce anything. This works because language models default toward statistically common outputs; naming the traps preemptively shifts probability mass away from them before generation begins.

The Naive Prompt

Write a cold email to a CFO at a mid-size logistics company 
pitching our AI-powered freight audit software. Keep it short 
and professional.

Why It Falls Short

“Short and professional” is instructions the model has seen ten million times, so it defaults to exactly what every other AI-generated cold email looks like: a subject line starting with “Streamline Your,” an opening compliment, three bullet points, and a calendar-link CTA. The result is technically competent but immediately recognizable as templated, which is precisely what kills cold email response rates. You asked for good output without telling it what “bad” looks like in this specific context.

The Improved Prompt

Write a cold email to a CFO at a mid-size logistics company 
pitching our AI-powered freight audit software. Keep it under 
120 words.

FAILURE MODES TO AVOID — and why they're tempting:
- Opening with a compliment or "I hope this finds you well": 
  feels polite but signals mass outreach immediately.
- Bullet-pointing our features: tempting because it looks 
  organized, but CFOs don't care about features; they care 
  about money recovered.
- Vague ROI claims like "significant savings": sounds 
  impressive but creates zero urgency because it's 
  unverifiable.
- Ending with "Would love to connect for a quick call": 
  overused to the point of invisibility.

Lead instead with a single specific financial hook (e.g., 
industry benchmark: logistics companies overbill 1.2–3% of 
freight spend). Close with one low-friction ask.

Why It Works

Naming each failure mode and its seductive logic forces the model to actively route around those patterns rather than fall into them by default — it’s the difference between “avoid clichés” (vague) and “here’s why clichés form and what they look like here” (actionable). The model now has a negative boundary around the output space, which tightens it toward genuinely differentiated copy. The financial hook instruction also gives it a concrete alternative to reach for, so it isn’t just constrained — it’s redirected.

When to Use This

High-stakes persuasion tasks (sales copy, pitch decks, grant proposals) where the model’s statistical average is also the reader’s pattern-match for “generic” — the exact thing that gets ignored.
Any domain with strong genre conventions the model will reflexively reproduce: legal disclaimers, performance reviews, apology emails, LinkedIn posts. If you’ve seen a hundred that look alike, the model has too, and it will write the hundred-and-first.
When iterating on outputs that feel “fine but flat” — if revision one came back technically correct but somehow lifeless, Failure Mode Injection on revision two is often faster than trying to describe the positive quality you want, which is frequently harder to articulate than the negative patterns you’re trying to escape.

Next edition of The Prompt Lab: another technique from the prompting toolbox. Catch up on the full archive at Stochastic Sandbox.