SuperAPI AI Directory

atomic-message

Office & productivity

View repository

Updated 10-15 21:40

About this skill

Atomic Message

Draft one outreach message at a time: a cold email, a LinkedIn note, or a follow-up. Generate a primary draft plus 1-2 variants, then lint against the craft rules before emitting.

This skill writes ONE message. It does NOT design multi-touch sequences, cadences, or step arcs, and it does not decide send timing. That is email-prompt-building's job. When a sequence needs each touch written, that skill calls this one per touch.

Inputs

Input Required Source
Signal yes the reason to reach out now: job change, hiring JD, event, post engagement, funding, or a specific observation. From upstream enrichment, the user, or research.
Persona yes the recipient's role cluster (e.g. founder, revops, sales leadership, IC). Shapes which pain and proof to lead with.
Channel yes one of: cold email, linkedin, follow-up. Sets length, format, casing, and CTA style.
Recipient context no name, company, role facts. Used to make the message relevant, never recited back.
Company context file yes voice, sender identity, what-we-do, positioning, proof library, CTA options, banned words. Read at runtime.

Workflow

Step 1: Read the context file

Read the company context file's Voice, What We Do, Proof, CTA, and banned-words sections. The skill supplies the craft (how to make an atomic message good); the context file supplies the company's voice and facts. If the context file's Voice section defines a per-channel spec, it overrides the generic channel defaults in references/craft.md.

Step 2: Pick the angle

From the signal + persona, choose exactly one of each:

  • the ONE thing to open on: the signal, or a sharp problem line that does not repeat the solution
  • the ONE capability framing relevant to that persona (from What We Do)
  • the ONE proof point relevant to that persona (from Proof; never stack more than one)
  • the ONE CTA (from the context file's CTA options; one per message, never stacked)

One signal, one framing, one proof, one ask. If you have three angles, that is three candidate messages, not one crammed message.

Step 3: Draft for the channel

Follow the channel's structure and length in references/craft.md, applying the context file's voice. Produce:

  • 1 primary draft
  • 1-2 variants that change the ANGLE (a different signal framing or a different proof), not just reworded synonyms

Step 4: Self-check (lint)

Run every draft through the self-check in references/craft.md before emitting. Read each line aloud in your head first. Fix anything it flags.

Step 5: Emit

Output the primary draft and variants as plain text (no markdown blockquotes). For cold email, include subject-line options. Label what each variant changes so the user can choose.

Reference

See references/craft.md for the universal copy mechanics, the per-channel structure (cold email / linkedin / follow-up), and the self-check lint.