Prompt Engineering for Recovery Email Copy

Generic AI-written recovery emails perform worse than the templated ones they replace. The difference between "LLMs are useless for this" and "LLMs 3x'd our recovery" is the prompt architecture, the guardrails, and the evaluation harness. All three, not any one.

# Structured inputs beat free-form prompts

Do not prompt with "write a recovery email." Prompt with a structured object. A minimal schema:

{
  "customer": { "first_name": "Anna", "signup_locale": "es-MX" },
  "product": { "name": "Netflix Standard", "brand_voice": "casual_confident" },
  "subscription": { "plan": "Standard", "price": "$15.49", "renewal_day": 12 },
  "failure": {
    "code": "insufficient_funds",
    "category": "soft_timing",
    "attempt_number": 1,
    "next_retry_at": "2026-07-17"
  },
  "cadence_step": { "channel": "email", "step": 1, "days_since_failure": 2 },
  "cta_url": "https://short.link/xyz",
  "constraints": {
    "max_subject_chars": 60,
    "max_body_words": 90,
    "tone": "friendly, non-shaming",
    "must_include": ["one CTA", "when the retry happens"],
    "must_avoid": ["urgency language", "the word delinquent"]
  }
}

Structured inputs let the model choose the right register — a es-MX locale + casual_confident brand voice + soft_timing failure produces a different email than en-US + formal + hard. The variance is what makes AI-written copy actually good.

# System prompt: role, rules, format

The system prompt should encode role, immutable rules, and output format. Not the input.

You are a bilingual copywriter for a subscription product. You write short,
friendly payment-recovery emails. You never shame the reader. You never use
urgency language ("act now", "final notice"). You never invent facts —
if a field is missing, omit that sentence.

Rules:
- Match the customer's signup_locale.
- Use the exact CTA URL provided; do not modify it.
- Keep subject under the max_subject_chars limit.
- Keep body under the max_body_words limit.
- Output valid JSON: { "subject": string, "preheader": string, "body_markdown": string }.

Two things this prompt does that most don't: it forbids invention ("if a field is missing, omit that sentence"), and it constrains output to a machine-parseable format. Both are essential.

# Guardrails that ship

Never send AI-generated copy directly to a customer. Between generation and send, run:

1. Length check — subject/body within limits.
2. Banned-phrase check — regex over "delinquent," "act now," "final notice," and 30 other landmine phrases you have discovered the hard way.
3. Fact check — the CTA URL matches exactly; the price matches exactly; the retry date matches exactly.
4. Language check — the output language matches the customer's signup locale.
5. Tone classifier — a small model or ruleset that flags shaming, urgency, or condescension.

Any failure sends the message back through generation once. Second failure falls back to a hand-written template. Do not send anything that failed twice.

# Evaluation harness

Prompt changes look like improvements 60% of the time and are actually regressions 30% of the time. Ship an evaluation harness:

• A frozen set of 200–500 real anonymized (customer, failure, cadence) tuples.
• Human-graded gold references for tone, clarity, and CTA presence.
• An LLM-as-judge scorer for tone match, factual accuracy, and constraint adherence.
• A/B production tests as the final arbiter.

New prompts must beat the current champion on the harness and in production. Regression in either kills the change. The engineering discipline is the difference between "we ship AI copy" and "we ship better copy than we did before AI."

# Locale is not translation

Do not generate an English email and translate. Generate directly in the target locale. Reasons:

• Idiom and register do not translate. "Hey" in English is "Hola" in Spanish but not "Bonjour" in French — the French equivalent of "Hey" is closer to "Salut."
• Cultural payment conventions differ. In Germany, receipts should reference the SEPA mandate ID. In Brazil, they should reference the Pix key or invoice number.
• Legal disclosures differ per market and cannot be translated on the fly.

The system prompt above accepts a locale field. Use it.

# Attribution windows and A/B

When testing AI copy against templates, hold the attribution window and channel constant. If you A/B new AI copy against a template that has been in market for a year, the template has an unfair advantage (it has already burned deliverability accumulated over time; the new one has not). Rotate both variants for at least 4 weeks before reading results.

# Fine-tuning: usually not

Fine-tuning a base model on your recovery emails is usually the wrong move. Reasons:

• The task is small (a few hundred lines of copy per language). Fine-tuning wastes signal.
• Base models updated by the vendor beat last-quarter's fine-tune in almost every generation.
• Guardrails and evaluation give you 90% of the benefit at 1% of the operational cost.

Fine-tune only if:

• Your brand voice is genuinely uncopyable by prompting (rare).
• You have >10k in-brand examples to train on.
• You have a lifecycle team that will maintain the tune across base-model upgrades.

# Cost budget

Each recovery email generation is 200–500 output tokens plus a small prompt. At current gateway rates that is fractions of a cent per email. Do not over-optimize model choice for cost — the recovery revenue per email dwarfs the generation cost by 4–6 orders of magnitude. Choose the model that produces the best copy under your guardrails; ignore the price.

# The RRLabs default

The Revenue Recovery Labs copy engine ships with structured input schemas per language, a system prompt hardened against 30+ landmine phrases, a guardrail pipeline with automatic fallback to templates, and a frozen evaluation harness that every prompt change must pass. AI-written copy converts 40–70% better than the templates it replaces — but only because none of the copy that reaches the customer is unreviewed AI output. It is AI output that survived the guardrails.