Write 5 different cold-email angles under 75 words for one persona

What this does

You give the AI one persona and one real fact about your offer. It returns five cold emails under 75 words, each built on a different angle (problem, peer, question, trigger, risk). You choose and edit one.

Who it’s for

SDRs and AEs writing outbound to a single persona. Not for blasting a purchased list. Not for situations where you have no honest reason to reach this person.

What you need

• Persona (title + industry)
• Your product in one sentence and one fact about it that you can back up
• One real reason to reach out (a signal you actually have), or ‘none’
• Your name and company, for the signature
• Tool: Any AI chat tool (for example ChatGPT, Claude, Gemini, or Grok)
• Cost: $0 with a free AI chat tier (limits change)
• Setup: 3 minutes (estimate)

Steps

1. Write down the persona and one fact about your offer that is true and provable (for example, something your customers have told you or your own pricing). If you don’t have one, write ‘none’.
2. Paste the prompt into your AI tool and fill the [VARIABLES].
3. Read all five. Pick the one that sounds most like how you talk, and cut anything you can’t back up.
4. Add your real signature block, your company’s postal address, and a plain opt-out line if your company or region requires it. The AI does not write these; it ends with a reminder to add them. Make sure no note to you ended up inside an email body.
5. Send it yourself from your normal email. Log which angle you used so you can compare replies later.

The prompt

You write cold emails for a salesperson. Write 5 versions of one first-touch email. Each version uses a DIFFERENT angle:
1. A problem the persona probably has
2. A peer comparison (only if I give you a real one in [PROOF]; otherwise skip and write a second problem angle)
3. A single curious question
4. A trigger (ONLY if I give you a real one in [SIGNAL]; otherwise write a question angle instead)
5. A risk of doing nothing

Persona: [TITLE + INDUSTRY]
What I sell, one sentence: [PRODUCT]
Proof I can back up (or "none"): [PROOF]
Real reason to reach out now (or "none"): [SIGNAL]
My name and company: [NAME, COMPANY]

NO-ADDITIONS RULE: Use only the facts I give you. Do not add facts, tips, statistics, scarcity or urgency, promises, references or testimonials, or the gender of any person that I did not supply. No fake familiarity ("I loved your recent post").

Rules:
- Each email is 75 words or fewer, including the greeting and sign-off.
- Plain words. No hype words (revolutionary, game-changing, unlock, supercharge).
- One clear, low-pressure ask per email.
- Subject line: 4 words or fewer, honest, no all-caps, no "Re:" or "Fwd:" tricks.
- If you skip an angle (because [PROOF] or [SIGNAL] is "none"), say so ONLY in that version's "Angle:" line. Never put a note to me inside a Subject or Body. Every Body must be ready to send.
- After each email, add one line: "Angle: ___ / Risk: ___" (one possible weakness).
- Do not write an opt-out line, postal address, or any statement about my company's identity.

Output format: Version 1 through 5, each with Subject, Body, and the Angle line. Then end with this line exactly:
"Compliance reminder for you (not part of any email): before sending, add your opt-out line and your company's postal address. I have not written them; you must supply them."

Sample input (fake data)

All names and companies are fictional. This is the input for run 4 (the output below), on the current prompt. Runs 5 and 6 used different inputs; see the test log.

Persona: Owner of a 15-person electrical contracting company
Product: Online quote-request form that sorts jobs by type and urgency
Proof: none
Signal: none
Name/company: Casey Morgan, SparkDesk (fictional)

Sample output

Real output from run 4 on 2026-10-14, using the current prompt, from an unspecified AI assistant (Grok Bot agent session; model name not shown). Copied as produced. Nothing was fixed after the fact; the problems we found in it are listed under “How it went”.

**Version 1**
Subject: Sorting quote requests
Body:
Hi,

Quote requests that arrive by phone, email, and text can be hard to sort by job type. Is that a problem in your shop?

SparkDesk is an online form that sorts requests by job type and urgency. Happy to send a short overview.

Casey Morgan, SparkDesk
Angle: Problem / Risk: assumes requests come in through several channels, which I wasn't told.

**Version 2**
Subject: Which jobs come first
Body:
Hi,

When several quote requests land at once, deciding which job to call back first takes time.

SparkDesk sorts requests by job type and urgency so the owner can see them in order. Would a short overview help?

Casey Morgan, SparkDesk
Angle: Second problem angle (no peer comparison was given in [PROOF]) / Risk: close to Version 1.

**Version 3**
Subject: Quick question
Body:
Hi,

How do you sort quote requests today: by job type, by urgency, or in the order they arrive?

If it's the last one, SparkDesk may be worth a look. If not, no problem.

Casey Morgan, SparkDesk
Angle: Curious question / Risk: vague subject line.

**Version 4**
Subject: How requests get sorted
Body:
Hi,

When a quote request comes in, who decides how urgent it is?

SparkDesk's form asks the customer for job type and urgency up front. Open to a short call?

Casey Morgan, SparkDesk
Angle: Question angle (no real reason to reach out now was given in [SIGNAL]) / Risk: nearly the same as Version 3.

**Version 5**
Subject: Cost of unsorted requests
Body:
Hi,

If quote requests aren't sorted, an urgent job can wait behind a routine one.

SparkDesk sorts by job type and urgency. Would it help to see how the form works?

Casey Morgan, SparkDesk
Angle: Risk of doing nothing / Risk: "an urgent job can wait" is a possibility, not something I know about your business.

Compliance reminder for you (not part of any email): before sending, add your opt-out line and your company's postal address. I have not written them; you must supply them.

Check before you send

• Word count is truly 75 or fewer (paste into a counter).
• No fact, number, or customer reference you can’t back up.
• Subject line is honest and matches the email.
• Signature includes what your company/region requires (sender identity, postal address, opt-out).
• The ask is something you’re willing to follow through on.

How it went

• Test date, tool, model: 2026-10-14. Unspecified AI assistant (Grok Bot agent session; the model name is not shown to the agent). One tool only.
• Runs logged: 6, on 6 different fake inputs (see the test log). Runs 1-3 used the original prompt. Runs 4-6 are a second round on new fake inputs after the prompt was fixed, as listed below. Every run was done by the same AI assistant session that wrote the fake input, and in round 2 the same session also edited the prompt, so these are not blind or independent runs. A script counted words and characters against the prompt’s limits. No second person re-ran it from the written steps.
• Why the label is “Tried in an AI tool,” not “tested”: we reserve “tested” for plays with at least 3 logged runs in at least 2 different AI tools, and we ran one tool. We also have no independent reproduction of the steps and no timed manual baseline. The runs are real, but this is one tool’s behavior, and 3 runs per round is too few to say a problem is gone.
• Time saved: Not measured. We did not time a manual baseline or an assisted run.
• What we changed after round 1: Added: no-additions rule; ‘If you skip an angle … say so ONLY in that version’s Angle line. Never put a note to me inside a Subject or Body.’; a closing compliance reminder (opt-out line, postal address) that the model must not write for the rep.
• Round 2 (current prompt, runs 4-6). What worked:
  • Rep-facing notes stayed out of the email bodies in all 15 emails (4 of 15 had them in the first round). Skipped angles were explained in the Angle line instead.
  • No peer counts, ‘most offices I talk to,’ invented features, ‘not selling anything,’ or ‘congratulations’ lines appeared this round (script scan plus a read).
  • The compliance reminder appeared at the end of all 3 outputs and nothing wrote an address or opt-out line. Word counts: bodies 24-47 (limit 75).
• Round 2: what failed or needed fixing:
  • Run 6 V2 subject ‘Booking link for recurring jobs’ is 5 words, over the 4-word limit. A script caught it; a reader might not.
  • Run 6 V3 ‘TidyLink might save a step’ implies a benefit nobody supplied. The model flagged it in the Risk line but left it in the body. The no-additions rule reduced invented claims but did not remove every one.
  • Several emails still assume how the prospect works today (for example ‘requests come in by phone, email, and text’). They were flagged in Risk lines, but the assumption is still in the body.
  • Fallback versions still resemble the angle they replaced (run 4 V2 and V4, run 5 V4), and the model said so in its Risk line.
  • The reminder is one line at the end, so it can be skipped if a rep copies only one email.
• Checked against “Check before you send”: Word counts passed except one subject. ‘No fact you can’t back up’ mostly passed; run 6 V3 needs a cut. Subjects were honest, but 3 of the 15 were a version of ‘Quick question’.
• Edge cases: Both ‘none’ (run 4) now works without notes in the body. Subject length still needs a count.
• Round 1 (original prompt, runs 1-3): what failed, kept for reference:
  • Rep-facing notes ended up inside the email body in 4 of 15 emails (run 1 V2 and V4, run 2 V4, run 3 V2), for example ‘No peer comparison was provided, so this is a second problem angle.’ The prompt says ‘say so’ but not where. If the rep doesn’t read closely, that line gets sent.
  • At least 6 of 15 emails went beyond the input: ‘Most HVAC offices I talk to’ (implies a sample), ‘leave a bad review’, an invented portal feature, ‘Not selling anything in this note’ (misleading), ‘Congratulations on the growth’, and ‘usually means rebuilding the schedule by hand’.
  • Fallback versions largely repeated the angle they replaced (run 1 V2 is close to V1, V4 close to V3).
  • None of the emails included a postal address or opt-out line. The prompt doesn’t ask for them; step 4 puts that on the rep.
• Not yet observed: how a different model or tool handles the same prompt (a second-tool run is still to do).

Watch-outs

Email: in the US, CAN-SPAM covers commercial email including B2B. Use your real identity, an honest subject line, a valid postal address, and honor opt-outs. EU/UK (GDPR/PECR) and Canada (CASL) can be stricter. Not legal advice. Don’t paste personal data about prospects you don’t have a reason to hold.

See the Safe Use page: AI drafts, you decide, you send.

Variations

• Manager version: ask for five versions the team can A/B test over two weeks, one angle per rep.
• Owner version: swap persona for ‘homeowner’ or ‘office manager’ and keep the ask to a 10-minute call.
• LinkedIn version: see ‘LinkedIn connect note + 2 follow-ups’.

Sources

None. Original wording. (Add any public framework or post used, with link and checked-on date.)

Retest

Re-check by: 2026-12-13 (60 days after the 2026-10-14 runs), or sooner when run in a second tool.

Test log

Changelog

• 2026-10-14: Draft written. Not run yet.
• 2026-10-14: Ran 3 times in one AI assistant on fake data. Filled in sample input and output, how it went, and the test log. Front matter set to runs_logged 3, last_tested 2026-10-14, retest_by 2026-12-13. The 2-tool bar for the word “tested” is not met, so the play is not labeled tested. The prompt text was not changed after the runs.
• 2026-10-14: Prompt revised after round 1. Added: no-additions rule; ‘If you skip an angle … say so ONLY in that version’s Angle line. Never put a note to me inside a Subject or Body.’; a closing compliance reminder (opt-out line, postal address) that the model must not write for the rep. Re-ran 3 fresh runs (4-6) on new fake inputs in the same AI assistant. Updated the sample input and output to run 4, the “How it went” section, and the test log (original runs 1-3 kept, marked). Front matter set to runs_logged 6, last_tested 2026-10-14, retest_by 2026-12-13. The 2-tool bar for the word “tested” is not met, so the play is not labeled tested.
• 2026-10-15: Published with the evidence label “Tried in an AI tool” (6 runs, 1 tool, fictional inputs). Earlier wording that implied more evidence than one tool’s runs was replaced with the label; the runs, results, and limits above are unchanged.