Agent-Skills.md

Agent Skills: Writing Copy

>

UncategorizedID: axiomantic/spellbook/writing-copy

Repository

axiomantic/spellbook
axiomanticLicense: MIT
105

Install this agent skill to your local

pnpm dlx add-skill https://github.com/axiomantic/spellbook/tree/HEAD/skills/writing-copy

Skill Files

Browse the full folder contents for writing-copy.

Download Skill

Loading file tree…

skills/writing-copy/SKILL.md

Skill Metadata

Name
writing-copy
Description
>

Writing Copy

Reasoning Schema

<analysis>Before drafting: whose thoughts these are, what the author actually observed, the audience and the piece's one point, and whether a personal voice skill is loaded.</analysis>

<reflection>After drafting: every AI tell in the catalog checked against the draft, every claim anchored to a number or an observable behavior, no argument invented that the author did not supply, and the piece handed back for a human edit pass.</reflection>

Invariant Principles

  1. The Argument Belongs to the Author: You sharpen and draft; you never supply the point. A piece whose thesis you invented is your writing wearing the author's name.
  2. Slop Is a Bounded Catalog, Not a Vibe: Every removal names the specific tell it removes. "This sounds like AI" is not a finding until it resolves to an entry you can point at.
  3. A Claim Without a Concrete Is Not a Claim: Attach a number or an observed behavior. To test a line that sounds good, flatten its cadence and read what remains; if the bare claim collapses into a generality, rhythm was carrying it and the line said nothing.
  4. Voice Is Borrowed, Slop Removal Is Owned: This skill is voice-neutral. On conflict the author's voice skill wins on register and phrasing; this skill wins on slop.
  5. Only Demonstrated Edits Become Rules: The catalog grows from real human edits the author made, never from taste you inferred on their behalf.

Overview

A voice-neutral guide for writing outward-facing prose that reads like a human wrote it, not an AI. It does two things: it runs the writing PROCESS (turn rough thoughts into a grounded draft), and it strips AI TELLS (the finite, nameable set of moves that make text read as generated).

It deliberately does NOT supply a person's identity. If the author has a personal voice skill (for example my-voice-skill), load that too: the voice skill owns who the author is (register, code-switch, signature phrasings), this skill owns process and slop removal. Core principle: AI slop is a bounded catalog of tells. Name them, cut them, and ground every claim in a concrete specific, and the prose reads human.

When to Use

  • Drafting or editing any outward-facing prose: newsletter, blog post, announcement, marketing copy, email, release notes, a doc's intro paragraph.
  • A draft "sounds like AI" and you cannot say why. The catalog below names why.
  • Turning a braindump or bullet notes into a finished piece.

When NOT to use:

  • Code, code review, or technical logic (the tells here are prose tells).
  • Reproducing a specific person's voice with no slop concern: load their voice skill.
  • Neutral machine-facing text where human texture is irrelevant.

The process

Writing well with AI is not one button. It is a loop that keeps the thoughts the author's and the voice the author's.

  1. Braindump first. Start from the author's raw notes: topic, audience, and loose bullets. If they are missing, ask for them. Do not invent the argument.
  2. Be a thought partner, not a ghostwriter. Before drafting, ask sharpening questions. Notice when the notes hide two or three different pieces. Reflect back the possible shapes; let the author pick. The main point is discovered here, not invented by you.
  3. Draft against known taste. Write the draft, then run the AI-tell catalog on your own output before showing it. Ground each claim in a concrete specific.
  4. Hand over for a human edit pass. The author decides what sounds right. Your draft is a starting point, not the finished piece.

The AI-tell catalog

Each entry: the tell, then a before -> after. These are voice-neutral; a personal voice skill may add more.

Vague, effort-theater, and jargon phrasings (say the concrete thing). The biggest tell, and the one humans rarely produce: filler that sounds like effort or a rating but names nothing. Cut "I went deep", "did a deep dive", "really dug into", "top box rating", "best-in-class", "move the needle", "at the end of the day", "when it comes to", "in terms of", "it's important to note", "a lot to unpack", "in today's world". Replace with the specific action, number, or behavior.

  • "We went deep on performance." -> "We cut p95 latency from 800ms to 210ms."
  • "It performs well." -> "It hallucinated once in about 200 runs, and obviously."

Overclaimed confidence (intensifier adverbs). Cut genuinely, truly, really, actually (for emphasis), very, quite, literally. They push a claim harder than the observation supports. Calibrated hedges that name real uncertainty ("I think", "probably", "seems like") are fine and stay.

  • "This is genuinely the right tool." -> "This feels like the right tool, because..."

Rhetorical compression moves. AI reaches for rhythm over information.

  • X-not-Y: "it isn't a rewrite, it's a refactor" -> "it's a refactor. No behavior changes." (Hot zone: opening sentences.)
  • Rule-of-three tricolons: "no install, no new tab, no new habit" -> one plain clause.
  • Quantified punchline: "three bullets instead of fifteen" -> plain description.
  • Comparative tail-end: cut "more X than anything else I use" from "than" onward.
  • Manufactured punchline ending: a closing line tighter and punchier than the paragraph ("The discipline isn't free.") -> end on the plain sentence before it.
  • Section-opener signpost: "A few things to flag:" -> open with the content.

Metaphor-loaded verbs and nouns. "becomes load-bearing", "users route around it", "unlock", "double down", "lean into", "supercharge". Use the plain verb.

  • "This becomes load-bearing." -> "This is now required."
  • "Users route around the feature." -> "Users ignore the feature."

Copula displacement. "serves as", "stands as", "acts as", "functions as" inflate a plain "is". Use "is" or the specific verb.

  • "The cache serves as the source of truth." -> "The cache is the source of truth."
  • "This acts as a fallback." -> "This is the fallback."

Formatting tells.

  • Em-dashes (the easiest tell to leave behind). Use parentheses, comma clauses, a hyphen with spaces ( - ), or two sentences.
  • Emoji-prefixed headers ("BREAKING"), over-bulleting that enumerates everything, "Sources:" citation blocks, relentless positivity. Cut or flatten to prose.

Ground every claim in the concrete

The positive rule behind half the catalog. When you assert something works or fails, attach a number or a specific behavior, not an adjective.

  • Not "it works pretty well" but "it has failed once in a couple hundred runs".
  • Not "it wasn't reliable" but "scheduled jobs failed and it hallucinated URLs". Prefer observable behavior over abstract characterization. A coined label is fine if you anchor it with concrete detail in the same sentence.

Emphasis-source check

When a line sounds good but says little, flatten the cadence: strip the rhythm and read the bare claim. If it collapses into a generality, the rhythm was carrying it, not the content. Rewrite around the concrete point. If it still lands flat, keep it.

Layering with a personal voice skill

This skill is neutral; a person's voice is not. If a voice skill exists for the author, load both:

  • Voice skill: identity, register, code-switch by audience, signature phrasings.
  • This skill: the process and the universal anti-slop catalog.

On conflict, the voice skill wins on voice and register; this skill wins on slop removal. A voice skill should inherit this catalog rather than restate it, and add only the person-specific tells and preferences on top.

The refinement loop (keep this guide alive)

The catalog gets better by learning from real human edits. After the author edits your draft:

  1. Put three things side by side: your draft, the author's edit, and this skill (or their voice skill).
  2. Classify each change: one-off wording, or a repeatable rule the guide is missing?
  3. For repeatable ones, propose a MINIMAL addition (a new tell, a tightened rule, a before -> after pair), grounded in the actual edit.
  4. The author accepts or rejects. Only write accepted changes. Naming what-not-to-do is how tacit taste becomes an explicit, reusable rule. Do not invent rules the author did not demonstrate.

Quick reference

| Tell | Fix | |---|---| | "I went deep", "top box rating", "move the needle" | Name the concrete action / number / behavior | | genuinely, truly, really, very | Cut the intensifier; keep calibrated hedges | | "isn't X, it's Y" | One plain statement | | rule-of-three, quantified punchline | One plain clause | | manufactured punchline ending | End on the plain sentence before | | "A few things to flag:" | Open with the content | | load-bearing, route around, unlock | Plain verb | | em-dash | Parens, hyphen with spaces, or split | | serves as / stands as / acts as | "is" or the specific verb |

Common mistakes

  • Deleting a punchy line but not replacing it. The strongest fix names what is actually happening (the mechanism), not just a shorter version of the slogan.
  • Over-hedging to sound human. Softening every claim is its own tell; cut intensifiers, do not add mush. Keep only hedges that name real uncertainty.
  • Restating a personal voice skill's rules here. This skill stays neutral; person- specific taste lives in the voice skill.
  • Inventing the argument. The thoughts are the author's; you sharpen and draft, you do not supply the point.