A reply is the only result a cold email can be judged by. Cold email writing is therefore the craft of removing everything that gives the reader a reason to close the tab, and five rules do most of that work.
Code: cold-email-coach, MIT licensed.
The skill lives as plain Markdown and a few references: 11 checks, a heuristic linter, and an installer that copies it into the agent directory.
Rule 1: the first line is evidence of attention
“Your post about 40% of new accounts never finishing setup stayed with me” cannot be written without reading something. “We help companies like yours” can be written without knowing who the reader is.
When no personalisation fact exists, the skill writes a placeholder instead of inventing a compliment. A bracketed [fact from their site] is honest, and it survives the review that catches an invented detail.
Rule 2: one ask, and it should be answerable in one word
A demo link, a calendar link and a question are three asks. Each one gives the reader a way to postpone all of them.
The strongest version of a cold email ends with permission to decline: “Is that still a problem, or is setup under control?” A person can answer that between two meetings, and that is where replies come from.
Rule 3: short beats clever
Fifty to one hundred twenty-five words. The subject line is short and neutral, because a clever subject announces a campaign.
A sentence that needs a second reading to parse is a sentence that will not be read twice. The skill cuts jargon for the same reason: a term the reader has to decode moves them out of the message.
Rule 4: no invented facts
No imaginary customers, no made-up metrics, no fake urgency. A real deadline is fine and belongs in the email; a manufactured one costs the trust the first line earned.
That rule applies to the model as much as to a copywriter. A skill that fills a gap with a plausible number produces an email that is embarrassing to send and worse to receive.
Rule 5: the score is a diagnostic
Eleven checks produce a score out of eleven, and a ten can still be irrelevant. The skill reports the score with that sentence attached, because a checklist that promises replies is selling something else.
A number that comes with a stated limit is more useful than a number presented as a verdict, and a team that reads the limit stops chasing a perfect score.
A scoring tool earns trust by refusing to overclaim. The same discipline appears in skills for ai agents: the file states what it cannot do, and the user knows where to look.
What the linter catches and what it misses
The Python linter counts words, reading grade, the you/I ratio, questions, asks, subject length, clichés, urgency words, capitals, exclamations, links and placeholders. It exits non-zero on a hard failure.
It cannot tell whether the observation is true or whether the offer fits. That is the model’s job and the sender’s, and the readme says so in one line rather than hiding it.
What it looks like in practice
Data, the offer and the problem go in one message; back come two or three short emails with a subject line and a word count. Then one is chosen and tuned for a specific person, because a template is recognisable by the third email anyway.
The check takes a minute: eleven points, a score, up to three fixes. A rewrite returns a short breakdown, a score and a ready text, and the breakdown explains why the first line drifted.
What this teaches about outreach
Write the email you would answer. The checklist is a formalisation of that question, and a team that runs it before a sequence goes out stops sending paragraphs about itself.
The skill is MIT, installs with one command, and works in any agent that accepts a system prompt or a skill folder.
Related: ai writing patterns for the phrases that make a personal email sound automated, and ai copywriting skill for the persuasion layer that should come after this one.