← Blog

Sell ·

Sales Message Review Prompts: the Pre-Send Check

Five prompts to review an outreach message before you send it: a brutal first pass, a rewrite that keeps your voice, a spam check and a follow-up writer.

By , Founder, Baljia AI

The short answer

You cannot read your own outreach message the way a stranger will, because you already know what you meant. The five prompts below give every message a second reader before it goes out: a brutal first pass, a claim audit, a rewrite that keeps your voice, a spam check, and a follow-up writer that references the first message without whining.

The difference from other prompt packs is that the honesty rules are written into the prompts themselves. Each one instructs the AI to mark what it cannot verify as unverified, to never invent a customer, number or fact to strengthen your message, and to stop and tell you when the fix requires real information it does not have. A reviewer that quietly upgrades your proof line is worse than no reviewer, because the claim goes out under your name.

We make Baljia, an AI employee for founders and small businesses; it shows up briefly at the end. The prompts run in any assistant.

Prompt 1: the brutal first pass

Review the outreach message below as the person receiving it, not as my
assistant.

Rules:
- Judge only what is written. Do not invent context that makes the message
  better or worse than it is.
- If a claim in the message cannot be checked from the message itself, mark
  it "unverified" rather than treating it as true.
- Do not rewrite anything yet. Review only.
- Be blunt. A polite review that misses problems costs me the reply.

The message:
[Paste the full message, with the subject line if it is an email.]

Who it goes to:
[One line: title, company type, and the evidence that they have the
problem.]

Score it pass or fail on each of these, quoting the text as proof:
1. Their problem first: does the first line say something about their
   situation, or about me?
2. Real proof: is there one line of evidence I have earned, and is it
   specific or vague? Mark it unverified if you cannot tell from the text.
3. A question, not a meeting ask: does it end with a question about their
   problem they can answer in one line, or does it push for a call?
4. Length: does it read as under 120 words? Flag it if it runs long; I
   will count the exact words myself.
5. The only test that matters: if this landed in your inbox from a
   stranger, would you reply? Yes or no, then one sentence on why.

The five checks compress the first-message rules from how to get customers from LinkedIn: open on their problem, add one line of earned proof, close with a question. Question 5 is deliberately unfair: the AI is not your prospect, but forcing a verdict surfaces the hesitations a checklist misses. Count the words yourself in your editor, because assistants miscount; the prompt asks for the judgment, the number is your job.

Prompt 2: the claim audit

Run this once per message, before the rewrite, because the rewrite must know which claims are allowed to survive.

List every claim in the message below and tell me which ones I can defend.

Rules:
- A claim is any statement about me, my product, my customers or their
  results. Extract each one exactly as written, and count them; do not
  merge two claims into one.
- Label each claim: defensible (I could show something real for it; say
  what), vague (cannot be checked either way), or risky (specific-sounding
  but unsupported by anything I gave you).
- Do not soften the labels, and never invent support for a claim. If you
  are unsure, the label is unsure, not defensible.

The message:
[Paste it.]

Delete or back every risky claim before going further. A claim you cannot defend costs you the exact moment it was supposed to win: the moment the reader checks.

Prompt 3: the rewrite that keeps your voice

Rewrite the message to fix only the failures from your review.

Rules:
- Keep my voice. Below are two things I actually wrote; match how they
  sound, not what they say. Keep my words and rhythm wherever they were
  not the problem.
- Do not add any fact, customer, number or claim that was not in my
  original. If the proof line was weak, tighten its wording; never upgrade
  it into a bigger claim.
- If the message fails because it needs real information I have not given
  you (a real observation about their company, real proof), stop and tell
  me what to go get. Do not paper over the gap with filler.
- Under 120 words. Shorter is better.

Writing that sounds like me:
[Paste two short things you wrote: a reply to a customer, a post, a work
message to a friend.]

My original message:
[Paste it again.]

Return the rewrite, then two lines on what you changed and what you
deliberately kept.

The voice samples are not decoration: "confident but friendly" produces the same smooth template everyone else sends, while two real samples give the model something concrete to match.

Prompt 4: the spam-trigger check

Read the message below first as a spam filter, then as a wary recipient.

Rules:
- Point only at concrete text. No generic advice.
- This is a read-only check; change nothing.
- Filtering rules are not public, so where you are unsure whether
  something triggers them, say unsure instead of stating it as fact. I do
  not want invented certainty.

The message:
[Paste the final version, with the subject line.]

Check for:
1. Template smell: lines that read like a hundred other people received
   them this week (flattery openers, fake familiarity, "quick question").
2. Pressure and hype: urgency, big promises, exclamation marks.
3. Mechanics: links in a first message, attachments, tracking-style URLs,
   odd formatting, all-capitals words.
4. Mismatch: a subject line that promises what the body does not deliver.

Return the specific phrases to change, why, and which items are known
problems versus your judgment calls.

Prompt 5: the follow-up writer

Write one follow-up to the message below. The first message got no reply.

Rules:
- Reference the first message in a single clause, without repeating its
  pitch and without complaint. Banned: "just bumping", "did you see",
  "circling back", "I know you're busy", and anything else that nudges or
  blames.
- Add one new thing that was not in message one, using only the material I
  give you below. If I gave you nothing worth sending, say so and stop;
  an honest "you have no reason to follow up yet" is a useful answer.
- Invent nothing: no facts, customers or numbers. Mark anything you are
  unsure of instead of asserting it.
- Under 80 words. End with a question they can answer in one line, not a
  meeting ask.

The first message:
[Paste it.]

The one new thing I can offer:
[One or two lines: something that changed, a thought, a resource you made.
Leave it empty if you have nothing.]

A follow-up without a new thing is a nudge, and the prompt is built to refuse to write one. If it stops on you, that is the pack working: go find the useful thing first.

How to read the output

A fail on question 5 means rewrite, whatever the other four said. Passing the checklist while losing the reader happens; the checklist diagnoses, the verdict decides.

"Unverified" is your job, not the AI's. The reviewer cannot check that your customer is real or your number is right. It can only refuse to assume. Everything it marks unverified goes on your list to confirm or cut.

When the rewrite prompt stops and asks for real information, do not negotiate. The gap it found is the reason the message was weak. Filler would only hide it from you, not from the reader.

Keep score across messages, not within one. Count unique people, never messages, and record the same three weekly numbers as how to get customers from LinkedIn: people who accepted, people who replied, meetings booked. One message's review tells you about that message; the weekly numbers tell you about the system.

What to do next

Review before sending, send at a human pace, one person at a time, and log what gets replies. The wider checklist for judging AI-drafted outreach is in how to review an AI-written sales message, and whether AI should ever press send for you has its own guide: is it safe to let AI send messages in your name?

Where Baljia fits: it is an AI employee for founders and small businesses that finds prospects and drafts LinkedIn and email outreach, with every message staying a draft until you approve it or switch sending on. LinkedIn campaigns come with the Growth and Scale plans. Details and limits are on the AI sales outreach page, plans on pricing.

Frequently asked questions

Should AI write the first draft or review my draft?

Review yours first, for a few weeks. Writing the draft yourself keeps the proof and the reason-to-contact honest, and the AI then catches what a stranger would stumble on. Once your messages pass routinely, drafting with AI from your own facts is fine, with the same review applied.

Why under 120 words?

Because the message has one job: to get a yes or no on the problem, and every sentence past that point gives the reader a reason to defer. The 120 figure is a working ceiling, not a magic number; good first messages land well under it.

Will cutting unverifiable claims make my message weaker?

It makes the message smaller and stronger: a claim you can defend survives the reader checking it, while one that collapses on inspection ends the conversation. If cutting the risky claims leaves nothing persuasive, you are missing real proof, and that is the thing to go build.

Does this work for LinkedIn messages and email both?

Yes, with one adjustment each way. LinkedIn first messages should run far shorter than the 120-word email ceiling, and email needs the subject-line mismatch check in prompt 4, which ordinary LinkedIn messages do not have, though InMail does. The five checks themselves do not change.

Run your next message through it

Take the message you were about to send today and give it the first pass. To see how Baljia works, start the 7-day free trial, no credit card, at https://baljia.ai/ and give it one focused task. LinkedIn campaigns come with Growth and Scale, not the trial.

Sources