Cold Message Writer

A Claude skill for writing cold outreach that actually gets replies — on email, LinkedIn, or X DMs.

Why this exists

Almost every cold message fails for the same reason: it's about the sender, and it reads like it was pasted into a list of 500 names. Title, company, CV, "we just shipped X" — the reader filters that out in two seconds because they get dozens of them a day.

The entire job of a cold message is to make one real person feel like you wrote it only for them. Get that right and the reply takes care of itself. This skill writes the opposite of the blast: short, specific, human, and easy to say yes to.

It's not a clever-words generator. It encodes the actual mechanics that get replies — making the reader feel chosen, opening with a hook about them not you, naming their objection before they can, making the ask weightless, proving demand with quiet specifics instead of adjectives, and sounding like a person typing between meetings rather than a polished template.

What's inside

SKILL.md                        the core instructions Claude loads when the skill triggers
references/templates.md         situation playbooks, ready-to-send templates,
                                the ghost follow-up sequence, annotated real examples
dist/cold-message-writer.zip    prebuilt zip for uploading to claude.ai

It covers five situations end to end — fundraising, hiring, partnerships/BD, press & podcasts, and job-seeking — each with its own hook, proof, and ask, plus a follow-up sequence for when people ghost (which is most of the time).

Install

This skill is plain instructions. No code runs. You can read every line before you install it.

claude.ai (Pro, Max, Team, or Enterprise)

1. Download dist/cold-message-writer.zip.
2. In claude.ai, open Settings → Capabilities → Skills. Make sure code execution is on.
3. Click Upload skill and select the zip.

That's it. Claude reads the skill automatically and uses it whenever you ask for cold outreach. Custom skills are per-user and don't sync across claude.ai, the API, and Claude Code, so install it on each surface you want it on.

Claude Code

Clone straight into your skills folder:

git clone https://github.com/felixleezd/cold-message-writer.git ~/.claude/skills/cold-message-writer

Claude Code discovers it on the next session. For a project-only install, clone into .claude/skills/ inside the project instead.

How to use it

Just ask in plain language. The skill triggers on its own — you don't have to name it.

• "write a cold email to a founder I'd like to work for at a seed-stage startup"
• "DM this investor about our round, here's the context…"
• "help me reach out to a podcast host about coming on the show"
• "follow up with someone who ghosted my first message"

By default you get one tight message. Ask for options and you get a few labeled variants that take genuinely different angles (lead with the mutual vs. lead with their post), not three tones of the same thing. Ask for follow-ups and you get the full ghost sequence.

Bring one true hook. The skill writes far stronger messages when you give it something real — a post they wrote, a launch they shipped, a mutual connection, a role they're hiring for. It will not fabricate a fake "I loved your post about X." Honest and specific beats polished and generic, every time.

The principles it runs on

If you read nothing else, read this. These are the moves the skill applies to every draft:

• Make them feel chosen. "I made a short list of people I wanted in early — you're on it." Specificity is what separates chosen from blasted.
• Open about them, not you. The first line is the notification preview. If it starts with "I," "we," or your company, it gets rewritten.
• Say their objection first. "You don't know me, feel free to ignore this — but I think it's actually for you." Naming it disarms it.
• Make the ask weightless. Not 30 minutes. A glance. "want the 90-second version? I'll drop it right here."
• Prove demand quietly. "showed five people Monday, four asked in" beats "everyone loves it." Ladder every claim to its hardest specific.
• Sound human, not polished. Lowercase, contractions, asymmetric rhythm, one real detail. Polish is the tell of a template — and a template is exactly what the reader filters out.

License

MIT — see LICENSE. Use it, fork it, adapt it, ship your own version.

Contributions welcome. If you have a cold message that worked, or a situation the playbooks don't cover yet, open a PR.