LeadHaste
IntermediateNo variables to fill — paste & go

Audit a batch of AI first lines before the campaign sends

Catches the failure modes that only show up across a batch — recycled constructions, swap-test failures, unverifiable claims — before prospects see them. You get a grade, pattern counts, individually flagged lines, and three paste-ready fixes for the generating prompt, turning QA from a vibe check into a repeatable gate.

The prompt
You are a cold email quality auditor who has reviewed personalization for campaigns sending millions of emails a month. You've seen every way AI first lines fail: invented facts, recycled sentence skeletons, compliments dressed as observations, lines that could go to any company in the industry. Teams call you before launch because you catch what row-by-row review misses — the patterns that only appear when you read fifty lines in a row.

I will give you a batch of first lines, each paired with the company it targets and, where available, the research data it was generated from.

Audit the batch and deliver:

1. BATCH GRADE (A–F) with a two-sentence justification.
2. PATTERN FINDINGS — repeated sentence constructions with counts ('14 of 50 open with Saw that...'), overused vocabulary, and any tonal drift.
3. FLAGGED LINES — every line that is: unverifiable against its research data, generic enough to pass the swap test (another company's name would fit), a disguised compliment, or factually suspicious. Quote each flagged line and state the failure in under 10 words.
4. FIX LIST — the three highest-impact changes to the generating prompt, written as instructions I can paste into it.

Be specific and unsparing: a polite audit that passes bad lines costs me replies and domain reputation. If the batch is genuinely strong, say so briefly and skip invented criticism.

Before you audit anything, interview me. Ask me these questions ONE AT A TIME, waiting for my answer each time:
1. Paste the batch — first lines with their target company names, plus the source research data if you have it.
2. Who is the campaign targeting, and what is the offer?
3. What prompt or process generated these lines?
4. What does your best-performing manual first line look like, if you have one?

Once you have my answers, deliver the audit in the four sections above. If the batch is over 100 lines, audit a representative 50 and say which you sampled.

How to use it

  1. 1

    Copy the prompt into Claude, ChatGPT, or any LLM.

  2. 2

    Export 50–100 first lines from Clay with company names and source-research columns, and paste the batch when the interview asks.

  3. 3

    Fix or discard every flagged line, and paste the FIX LIST instructions into your generating prompt.

  4. 4

    Regenerate, re-audit, and only launch a batch that grades B or higher — then re-run this audit on a sample every few sends.

Best practices

  • Always include the source research data — without it the auditor can only check style, and unverifiable-fact detection is the highest-value check.

  • Audit a random sample, not the top of the table — Clay tables are often sorted in ways that cluster your best-enriched rows first.

  • Track your batch grades over time; drift downward usually means a data source degraded, not that the writing prompt changed.

Example: what this looks like in practice

Before launching a 4,000-send campaign, a growth lead exports 60 first lines with their scraped-research columns and runs this audit. The batch grades C+: 19 lines share a 'While most X, you Y' skeleton, 6 fail the swap test, and 3 reference a customer count the auditor can't find in the source data — one scrape had captured a competitor's stats page. The FIX LIST adds a construction ban, a specificity rule, and a source-fact requirement to her generating prompt. The regenerated batch grades A-, and the campaign launches at a 5.2% reply rate instead of shipping 4,000 copies of a detectable template.

Best fit

Roles
SDR / BDRRevOpsFounder / CEO
Company size
Startup (1–10)SMB (11–50)Mid-market (51–500)
Audience
B2B
Industries
Any industry
Works with
Any LLM
Difficulty
Intermediate

This prompt is one gear in a bigger machine. We orchestrate 20+ tools into outbound systems our clients own — and guarantee the results.

Apply for a Pilot Spot →
Prompt FAQ

Frequently asked questions

Read them in bulk, not one at a time — quality problems in AI personalization are batch-level: repeated skeletons, interchangeable lines, drifting tone. A structured audit with pattern counts and a swap test catches in minutes what row-by-row review never sees. Grade B or better with zero unverifiable facts is a reasonable launch bar.