personalisation-line

---
name: personalisation-line
description: One personalised first line that makes templated cold email read as hand-written; for volume outreach to researched prospects.
---

# Get templated outreach answered like hand-written mail

Real-signal first lines, ready to send: 50 to 250% more replies than merge-field templates (Lavender, 231,818 emails).

## Claude does

1. Ask, one sentence each: what they sell, who to, the ask, the tone. Per prospect: name, company, raw material (About page, posts, news, job adverts). Guess nothing.
2. Research by Lavender's 5x5x5: five minutes, five sources, stop.
3. Rank facts: Level 1 random, 2 company, 3 recent, 4 recent and tied to the user's problem. Keep the highest; below 3, skip. Situation signals (hiring, funding, site change) beat compliments.
4. Decide Y, in the user's words, for "I saw X, and I believe you are dealing with Y" (Lavender). "So what?" test: X must point at Y, else back to 3.
5. Record: lead_id, name, company, signal_type (post, news, hiring, tech, profile, mutual), verbatim evidence (exact words or dated change), implication, status ok or skip.

## Then Codex does

Codex fits each signal and implication into a situation line plus a second option under hard caps, varying structure, and writes JSON and CSV: bulk work, nothing to decide.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Fill in: {WE_SELL}, {ICP}, {OFFER}, {TONE}, {SIGNALS} (step 5 records), {OUTPUT_FOLDER}.

## Claude checks

1. Same lead_ids as {SIGNALS}, same order, skips kept.
2. Deletion test: cover the line, read the rest; still fine means decorative, fail.
3. Every fact traces to the verbatim evidence; one invention fails the row.
4. Under 25 words, their situation first, no sender or product, clean hand-off.
5. No banned opener, attention hack, AI tell or repeated shape.
6. Five random rows as the recipient: obvious why this person wrote, not creepy.

Only failing rows go back, with reasons; twice failed becomes skip.

## Rules

- Personal means specific, recent, about them: something you could only know by looking.
- Post: three to six of their words, quoted. Change: dated. Profile detail: unusual. Connection: real.
- Never "I see you work at X", "Impressive background", the headline, {{FirstName}} and {{Company}} (Level 1, 5% lift), "I hope this email finds you well", "Allow me to introduce myself", "My name is", "I wanted to reach out" or "I saw your LinkedIn profile".
- No praise, attention hack or product opener: their world first, you over we; generic personalisation is worse than none.
- One signal, one detail; over-personalising reads creepy and blurs the ask.
- Public information only; thin data means skip, never a guess; the line goes out in the user's name. One sentence under 25 words; anyone-else-unchanged means rewrite.
- No leverage, synergy, ecosystem, em dashes or repeated patterns.
- Line only, never the whole email, three fields maximum; 25 of 100 points, no rescue for a weak list or offer.

## If Codex is not installed

Claude does the Codex section itself and says so.

## Built from

- coreyhaines31/marketingskills, https://github.com/coreyhaines31/marketingskills/blob/main/skills/cold-email/SKILL.md, 46,908 stars (api.github.com, 5 Sep 2026).
- alirezarezvani/claude-skills, https://alirezarezvani.github.io/claude-skills/skills/marketing-skill/cold-email/, 25,542 stars.
- growthenginenowoslawski/coldoutboundskills, https://github.com/growthenginenowoslawski/coldoutboundskills/blob/main/skills/campaign-copywriting/SKILL.md, 680 stars, 1,000+ real campaigns.
- Clay blog, https://www.clay.com/blog/ai-email-personalization-examples, official blog, 24 prompts.
- Instantly, https://instantly.ai/prompt/personalized-cold-email-first-line-generator-sales, 5,465 recorded runs.
- Lavender, https://lavender.ai/blog/how-to-build-a-cold-email-personalization-process, 231,818-email benchmark.

# Get templated outreach answered like hand-written mail - Codex task

Everything you need is below; you have no memory of the conversation behind it. Do not browse or ask.

## You are given

- {WE_SELL}: what the sender sells; context only, never in a line.
- {ICP}: who they sell to.
- {OFFER}: the ask the email makes.
- {TONE}: casual, formal or peer-to-peer.
- {SIGNALS}: JSON array, one object per prospect: lead_id, name, company, signal_type (post, news, hiring, tech, profile, mutual), signal (verbatim evidence: exact words or a dated change), implication (the problem it points at, in the sender's words), status (ok or skip).
- {OUTPUT_FOLDER}: where to write.

## Produce

In {OUTPUT_FOLDER}:

1. personalisation-lines.json: same lead_ids, same order, each with lead_id, name, company, source (signal_type unchanged), status, situation_line, second_option.
2. personalisation-lines.csv: same rows, header lead_id,name,company,source,status,situation_line,second_option; text fields double-quoted, inner quotes doubled.

Per ok prospect:

- situation_line: one sentence stating their exact situation (the signal) and bridging to the implication. Complete "I saw X, which usually means Y" with keywords from signal and implication only; vary the frame, never the facts. Under 25 words, under 15 better. Opens on them. Ends so the sender's next sentence follows without a jolt.
- second_option: same caps, a different angle on the same signal (a question exposing the gap, or the implication first), not a rewording.

Per skip prospect, or empty signal or implication: status skip, both text fields "<fields cannot be generated - skip>".

## Rules

1. Never fabricate: only facts in signal and implication; thin data means skip, not a filled gap.
2. Post lines quote three to six of their exact words; change lines carry the date; profile and mutual lines name the specific detail.
3. Never "I hope this email finds you well", "Allow me to introduce myself", "My name is", "I wanted to reach out", "I saw your LinkedIn profile", "I see you work at X" or "Impressive background". Never mention {WE_SELL} or any product.
4. No compliment unconnected to the implication.
5. No leverage, synergy or ecosystem. No em dashes; full stops or commas.
6. Vary the opening word and structure across prospects; no repeated pattern.
7. One signal per line; never combine two.
8. Could it go to anyone else unchanged? Rewrite until not.
9. Match {TONE}. British English.

## Return

Print, in order:

1. `rows: N, ok: N, skip: N` and both file paths.
2. `longest line: N words`.
3. First three words of every ok situation_line, one per row, to expose repeats.
4. Five random rows in full, for spot-check.
5. Any lead_id set to skip that arrived ok, with the reason.
6. The complete JSON.

Then print DONE.